Credit Lines as Insurance: Evidence from Bangladeshbarrett.dyson.cornell.edu/NEUDC/paper_447.pdf ·...

Credit Lines as Insurance: Evidence from

Bangladesh

Gregory Lane∗

October 1, 2018

[JOB MARKET PAPER](Click here for the latest version)

Abstract

When insurance markets are absent, theory suggests that households can use credit lines to insure them-selves against adverse income shocks. However, in many developing countries access to credit in the aftermathof shocks is scarce as negatively affected households are frequently denied loans. In this paper I test whethera new financial product that offers guaranteed credit access after a shock, allows household to insure them-selves against risk. To this end, I run a large scale RCT in Bangladesh with one of the country’s largestmicrocredit institutions. Microfinance clients were randomly pre-approved for loans that are made avail-able in the event of local flooding. I show that this unique type of microcredit improves household welfarethrough two channels: an ex-ante insurance effect where households increase investment in risky productionand an ex-post effect where households are better able to maintain consumption and asset levels after ashock. I also document that households value this product, taking costly action to preserve their guaranteedaccess. Importantly, extension of this additional credit improves loan repayment rates and MFI profitability,suggesting that this product can be sustainably extended to households already connected to microcreditnetworks.

∗Department of Agricultural and Resource Economics, UC Berkeley. Email: [email protected]. Igratefully acknowledge the financial support of BASIS and Feed the Future. I would like to thank my partnerBRAC for their collaboration, and in particular Hitoishi Chakma, Monirul Hoque, and Faiza Farah Tuba for theirinvaluable assistance. I thank Erin Kelley, John Loeser, and Edward Rubin for their assistance and Alain de Janvry,Jeremy Magruder, Aprajit Mahajan, Kyle Emerick, and Edward Miguel for their advice. Finally, I thank my advisorElisabeth Sadoulet for her incredible support and guidance throughout. All errors are my own.

https://www.dropbox.com/s/aueul4qqabw7vfo/JMP_draft.pdf?dl=0

mailto:[email protected]

Job Market Paper Gregory Lane

1 Introduction

Poor households throughout the developing world struggle with income risk. They primarily rely on

agriculture and small business profits, which are vulnerable to shocks including harvest failure after

extreme weather events, price volatility, and a sudden loss in the family (Dercon, 2002). Research

has demonstrated that these high levels of income variability prevent households from accumulat-

ing wealth and exiting poverty. Moreover, the set of risk coping and mitigation strategies that

are available to households can often leave them worse off in the long run. Many turn to low-risk

production technologies and under invest in inputs, which negatively affects their future returns

(Donovan, 2016). Others are often forced to implement damaging coping strategies such as lower-

ing food consumption, selling productive assets, and reducing health and educational investments

(Hoddinott, 2006; Janzen and Carter, 2018). Traditional insurance markets are designed to help

households cope with these risks, but they are often absent or incomplete in developing countries

because of moral hazard and adverse selection, while alternative tools such as index insurance have

been hampered by low demand (Jensen and Barrett, 2017; Cole and Xiong, 2017).

Theory suggests that a realistic alternative to these tools is to provide households with a credit

line so they can self-insure. Credit and savings models have long highlighted the precautionary

value of credit access, which can serve to insure households against income fluctuations (Deaton,

1991, 1992). However, there is little empirical evidence demonstrating that households can in fact

use credit access in this way. What evidence does exists comes from developed countries (Gross and

Souleles, 2002), even though the benefits of credit access will likely be larger in developing countries

where insurance markets are lacking. In developing countries, the largest providers of formal

credit to the poor are microfinance institutions (MFI), which severely curtail credit access in the

aftermath of large aggregate income shocks (Demont, 2014). Their requirements that households

be financially evaluated at the time of loan disbursal, and that households cannot borrow if they

have any outstanding debt, severely restricts households’ use of credit access as a buffer against

risk in developing countries.1

This paper provides the first empirical evidence that guaranteed access to credit after negative

shocks increases productive investment, improves households welfare, and ultimately, is profitable

for the MFI. To this end, I partner with BRAC, a large MFI in Bangladesh, and extend a guaranteed

credit line to poor, rural households. The product, marketed as an Emergency Loan, is a pre-

approved loan that is made available to clients when an aggregate local shock (in this case a flood)

occurs. I randomize the availability of the Emergency loan across 200 rural BRAC microfinance

branches serving over 300,000 clients with over one million loans during the study period. Clients

in the 100 treatment branches were informed, before the beginning of the planting season, that

they were pre-approved to take the new loan product should a flood shock occur in their area.2

1Many MFIs have a dual mission of profit making and increasing social welfare which theoretically makes themmore likely to extend credit to households after an income shock. However, the field staff responsible for approvingindividual loans are almost always evaluated primarily on repayment metrics, making them likely to avoid lending torisky households.

2The Emergency Loan was only offered to approximately 40% of clients within each treatment brach based on an

1


Control branches continued their normal microfinance operations. Loans were then extended upon

request by eligible households after a flood had occurred (and been externally validated).

The experiment documents four primary results. First, I find that households value access to

guaranteed credit and respond as theory would predict. Indeed, some households are willing to

forgo credit in the pre-period in order to preserve access to the state contingent Emergency Loan,

suggesting that at least a subset of clients value the precautionary benefits of credit access. Rough

estimates suggest that these households value credit access after a shock at least 1.8 times more

than credit access in the pre-period.

Second, I find that informing households that they are pre-approved for credit in the event of

a flood is associated with a significant rise in risky investments. Treated households increase the

amount of land dedicated to agricultural cultivation by 15% and increase non-agriculture business

investments. Both of these effects are concentrated among the most risk averse households. These

findings suggest that households view guaranteed liquidity access as reducing their exposure to flood

risk, and respond by increasing their investment in riskier, potentially more profitable investments.

Third, I document that emergency credit, unlike many other microcredit products, improves

household welfare outcomes. When there is no flood, the larger ex-ante investments translate into

higher revenues. When flooding does occur, households are better able to maintain consumption

and asset levels. Furthermore, we find that the most severely affected households were the most

likely to use this additional liquidity. This means that the largest gains associated with guaranteed

credit could be concentrated among those who need it the most.

Finally, I find that extending guaranteed credit to clients in the aftermath of shocks does not

harm (and marginally improves) overall MFI performance. Borrowers with access to the Emergency

Loan improve their overall repayment rate, driven by improvements in repayment rates after a flood

shock. Overall, there is suggestive evidence that branch profits weakly increases, with the largest

increases in profits coming from “marginal” clients. This result is encouraging for MFIs that

have traditionally withheld credit in the aftermath of aggregate shocks. Nevertheless, it is worth

highlighting that these results may not generalize to contexts where repayments rates are low to

begin with.

The provision of guaranteed credit lines combines aspects of traditional microcredit and insur-

ance products, both of which have been extensively studied in developing countries. The provision

of traditional (loss-indemnity) insurance is almost completely absent among low-income households

due to high administrative costs, adverse selection, and moral hazard (Jensen and Barrett, 2017).

In recent years, index insurance has been promoted as a viable alternative. By linking payouts to

easily measurable and exogenous indices such as rainfall, index insurance removes moral hazard con-

cerns and reduces the need to collect additional data on household specific losses. Index-insurance

has been found to generate positive results by inducing more investment in agricultural production

and reducing the sale of assets after shocks (Karlan et al., 2014; Janzen and Carter, 2018). Despite

these benefits, demand for index insurance remains very low across many developing countries when

individual credit score, see section 2 for more details.

2


offered without heavy subsidies (Cole and Xiong, 2017). Low demand appears to be linked to the

requirement that insurance payments be collected ex-ante, which can be difficult for households

that are 1) credit constrained, 2) present-bias, 3) face basis risk that the index will not correspond

to their own person shock, and 4) lack trust in their insurers ability to pay-out when the time comes

(Cole et al., 2013; Clarke, 2016). In recent work, Serfilippi, Carter, and Guirkinger (2018) show that

preferences for certainty drive down demand for insurance contracts where premiums are always

paid but payouts are uncertain. In some contexts low demand can be overcome by allowing the

upfront insurance premium to be paid after harvest. However, this solution is only feasible when

there is the possibility of an interlinked transaction. This can take the form of a monopsony buyer

that can credibly (and cheaply) collect payment from farmers after the fact, such as in Casaburi

and Willis (2018) or by tying insurance payments to credit contracts as in McIntosh, Sarris, and

Papadopoulos (2013).

This research demonstrates that emergency credit can function as a viable alternative to in-

surance products with several key advantages. Specifically, the Emergency Loan overcomes the

challenges associated with the timing of insurance payments while maintaining many of the positive

features associated with index insurance. Like index-insurance, the availability of the additional

credit is contingent on an exogenous indicator (floodwater height) to avoid high administrative

costs and moral hazard. However, unlike index-insurance, no purchase (or any binding decision) is

required by the household during the planting season (this is similar to the innovation explore in

Casaburi and Willis (2018)). Providing coverage under a guaranteed credit scheme simply requires

notifying a household that they are eligible for the product. As long as households understand the

offer, and trust that it will be executed if needed, the household is “treated”. This feature ensures

that credit constrained or present biased households that stand to benefit from the product will

not be deterred from adopting it. As a result, guaranteed credit lines have the ability to provide

coverage to a large number of households that might not otherwise choose to purchase insurance.

Critically, households can benefit from the security of the credit line even if they choose not to

take a loan after a shock. This arises because the decision to take credit is postponed until after

uncertainty has resolved, which means households can opt in or out depending on realized damages

from the shock and any alternatives that may be available. Indeed, I see in my experiment that

many households increase ex-ante investment, suggesting a reduction in perceived risk, even though

ex-post take-up of the Emergency Loan is low.

There are, however, several limitations associated with using guaranteed credit as a risk man-

agement tool. Like insurance, households may be reluctant to rely on the product in times of need

if they are concerned about default by the provider (this is mitigated in this context by working

with BRAC Bangladesh, a well established and trusted MFI in the region). Unlike insurance, the

sequence of shocks can have an impact on the usefulness of credit for income smoothing. If a house-

hold experiences multiple successive shocks under a guaranteed credit scheme, they may accumulate

excess debt or exhaust their available credit line.3 Finally, extending credit to households after a

3With insurance a household that experiences several shocks in a row will simply receive the fixed insurance payout

3


shock is inherently risky for MFIs. While I find good repayment rates in this setting, if repayment

rates are lower elsewhere, providing guaranteed credit may not be sustainable from the lender’s

perspective. It follows that while guaranteed credit provides clear advantages to some households,

it may not be a panacea.

This research also contributes to the large literature on microcredit. Developed in Bangladesh

in the 1970s, microcredit institutions have since rapidly expanded, reaching over 137 million house-

holds worldwide Maes and Reed (2011). Unfortunately, despite this extensive growth and early

enthusiasm for microcredit4, the majority of research documents only modest impacts on house-

holds’ well-being (Karlan and Zinman, 2011; Angelucci, Karlan, and Zinman, 2015; Banerjee et al.,

2015; Banerjee, Karlan, and Zinman, 2015). This may be partly attributable to the fact that

microcredit only solves the problem of credit access, without remedying the underlying risks that

prevent households from optimally investing (Karlan et al., 2014). Indeed, early microcredit prod-

ucts typically featured group lending with joint liability and frequent, rigid repayment schedules

designed to overcome high transaction costs and asymmetric information, but that come at the cost

of making repayment difficult for those with uncertain income (Karlan, 2014). In response to these

results, a line of research has focused on easing these constraints by matching repayment schedules

to borrowers’ cash flows. Field and Pande (2010) and Field et al. (2013) show that reducing pay-

ment frequency and delaying the start of installment payments reduce borrower transaction costs

and encourage greater investments and profits. Similarly, Beaman et al. (2014) study agricultural

loans that allow repayments to come in a lump sum after harvest, and find higher investments in

the planting season and Barboni (2017) shows that more productive borrowers opt into flexible

repayment contracts even when they are more expensive. This paper builds on these results by

showing that credit products that increase the flexibility of households’ access to credit (not just

repayment flexibility), can lead to important improvements in outcomes.

Lastly, additional research has focused on understanding how new credit products affect MFI

profits. Field et al. (2013) develop a structural model to show that longer grace periods are not

sustainable for MFIs due to adverse selection and moral hazard concerns. In contrast, Barboni

(2017) use theory and lab-in-the-field experiments to show that offering flexible repayment schedules

could increase profits for lender. An advantage of this relatively large experiment is that allows

an empirical examination of the effects of this new product on overall MFI profitability, which

is difficult in settings where risk-averse MFIs are hesitant to experiment. In this setting, I find

significant positive effects on MFI profits, with significant heterogeneity among borrowers.

The rest of the paper is organized as follows. Section 2 describes the context of the experiment

and describes the new credit product in detail. Section 3 lays out a theoretical framework which

provides predictions. Section 4 describes the main research design and execution of the experiment

and section 5 describes the data used in the analysis. Finally, section 6 present the results of the

experiment and section 7 concludes.

each period (provided they purchased the product).4In 2006, Mohammad Yunus and the Grameen Bank were awarded the Nobel Prize for Peace.

4


2 Context and Product Description

Bangladesh and Income risk

This project takes place in Bangladesh, a country with over 165 million people that is covered by the

Bengal delta (a confluence of the Ganges, the Brahmaputra and the Megna rivers). Approximately

70 percent of Bangladesh’s population lives in rural areas and more than 80 percent of rural

households rely on agriculture for some part of their income (World Bank, 2016). While the

country’s economy has grown rapidly in recent years, GDP per capita still stands at $2,363 and

approximately 43% of the population earns less than $1.25 per day (UNDP, 2015).

Many types of extreme weather events are frequent, and are projected to worsen with the advent

of climate change. Approximately 80% of the country is located on floodplains, and floods occur

yearly with varying degrees of severity (Brammer, 1990). Moreover recent projections estimate that

flood areas could increase by as much as 29% in Bangladesh (World Bank, 2016). Therefore, the

experiment focuses on flood risk over other shocks because of the high frequency of flooding across

the entire country. As such, the randomized control trial was conducted in areas located close to

the major rivers where frequent flooding occurs.

I work with a subset of households that are active microfinance borrowers. These households

are primarily engaged in agriculture: 50% grow their own crops and 22% work as day laborers. This

group is also active in starting their own businesses (27% reported owning a small shop). Education

is low in these areas, and approximately two thirds of the sample have less than a primary school

education.

BRAC Microcredit

BRAC was founded in 1972 in Bangladesh and is currently one of the largest NGOs in the world.

Their microfinance operations began in 1974 and have expanded to serve the entire country. They

operate over 2000 branches where each branch serves anywhere from 20 to 60 village organizations

(VO’s). These organizations are designed to facilitate coordinated activities between borrowers at

the village level, including the distribution of information about new micro-finance products and

creating a convenient space for BRAC loan officers to collect loan payments and instill some social

pressure on borrowers to make their loan payments. VO meetings occur either weekly or monthly

depending on branch policy. At each meeting the loan officer collects the scheduled loan repayments

from each active borrower and answers enquiries about desired new loans from members without

existing debt.

The most common loan provided by BRAC is called the Dabi loan, which is only given to

women and is targeted at poor households5. Dabi loans are typically small in value (approximately

15,000 taka ($187)), and are required to be repaid within a year. Microfinance interest rates are

regulated in Bangladesh and BRAC charges 25% interest on the Dabi loan, which is near the legal

5While the Dabi loans are given only to women, it is common that these loans are used for broader householdinvestments such as agriculture or a business that is run by the husband of the official borrower.

5


maximum and similar to other MFI’s. During the repayment period, borrowers are not allowed to

apply for any other BRAC loans, and are discouraged from taking any additional loans from other

microfinance institutions or local money lenders. There is, however, one exception. Clients who

make every loan payment on-time for the first six months of their loan cycle are eligible to take a

top-up loan called the “Good Loan”. The Good Loan is capped at 50% of the principal amount

of the currently held loan. The offer is only available for two months after they become eligible at

the 6 month mark of their current loan cycle.6 In every other respect, Good Loans are identical

to normal Dabi loans, with the same 25% interest rate and one year repayment timeline. Taking

the Good Loan does not delay the normal loan cycle, and the client can take another normal Dabi

loan as soon as they have repaid their old one.

Product Description

The Emergency Loan was designed together with BRAC to improve its utility for borrowers exposed

to flood risk while also limiting BRAC’s exposure to risky loans. Clients were eligible to access

the Emergency Loan provided they had a credit score above a fixed threshold. The credit score

was created specifically for this product, and was calculated from each borrower’s past repayment

behavior on previously held BRAC loans. Specifically, the score was based on four metrics: past

percentage of missed payments, average percent behind on loan payments, maximum percent behind

on any loan, and the number of months as an active BRAC microfinance member. Each variable

received a linear weight determined by a regression of these variables on a binary indicator for

loan default. This weighted sum was then normalized to a 0-100 scale. The variables themselves

were chosen based on several criteria, including a) ease of calculation due to record keeping and

computation limitations, b) relevance for predicting future default, and c) ease of explanation for

transparency.7 The threshold was set so that approximately 40% of borrowers were eligible at any

given branch (77 out of a maximum score of 100). It is worth highlighting that targeting based

on credit score does not select richer households over poorer ones. Table 1 examines differences in

observables between the eligible and ineligible borrowers. The two groups look fairly similarly, but

differ along a few dimensions. Eligible borrowers have slightly less annual income, they are a few

years older, have fewer years of education, and own more livestock and savings.

Client eligibility was assessed for every borrower in April, just before planting of the Aman

season and several months before the flooding season. Borrowers could retain access to their Emer-

gency Loan eligibility for the duration of the Aman cropping season regardless of their repayment

behavior in the interim. Eligible clients were guaranteed to be able to borrow up to 50% of the total

principal amount of their last regularly approved loan. For example, a borrower who took a 10,000

6Good Loans are also subject to the Loan Officer and Branch Manager approval (i.e. they can be denied even ifthe borrower is technically eligible)

7To determine relevance for predicting default, the complete set of possible variables was assessed in two historicaltraining samples and then confirmed using more recent data. Only variables that were consistently predictive werekept in the final credit score. Additionally, linear regression was used rather than more complex techniques such asmachine learning due to the desire to make the credit score transparent, and easily adjustable in the future.

6


taka loan ($125) in May from BRAC was guaranteed to borrow up to 5,000 taka ($63) should a

flood occur regardless of their existing loan balance. No further evaluations of the client’s ability

to repay, or any other checks, were conducted before disbursing the Emergency Loan. Emergency

Loans were then made available to eligible clients if flooding occurred. This was validated in two

ways. First, the river gauge associated with the branch area had to be reporting water level above

the pre-determined danger level for at least one day.8 Second, a non-microfinance BRAC employee

had to confirm that at least 20% of the branch service area had experienced flooding.9 On a case by

case basis, loans were also made available if the local Branch Manager notified BRAC headquarters

of local floods and this report was confirmed by the BRAC employee (even if the matched river

water gauge had not passed the official danger level).

Two additional features of the Emergency Loan are important to review. First, the eligibility

list created by the credit score was provided directly to branch managers who could veto up to 10%

of the names on the list based on their private knowledge of a borrower’s credit worthiness.10 The

final list was then shared with BRAC headquarters. These steps were put in place to minimize the

risk that BRAC would lend to borrowers that would fail to repay the loan.11 For the purposes of

the experimental results, I do not consider Branch Manager vetos and include all clients that were

determined to be eligible based on the credit score alone.

Second, it is important to note how the Emergency Loan interacts with the existing Good Loan

product. The Good Loan product differs from the Emergency Loan in the timing that is made

available to clients (6-8 months into their normal Dabi rather than post-flood), and in how it is

disbursed (by asking branch managers who can deny the request, instead of pre-approval based on

credit scores). Looking through historical data this means that the Good Loans were much less

likely to be disbursed in the aftermath of aggregate income shocks because most borrowers were

either not in their 6-8 month timeframe and branch managers did not want to approve additional

top-up loans.

Clients in the sample could be eligible for the Good Loan or the Emergency Loan, both, or

neither. However, borrowers were informed that they could not have both a Good Loan and an

Emergency Loan together — if they take a Good Loan they lose future eligibility for the Emergency

Loan should a flood occur.12 This limitation was introduced because of BRAC’s concerns that

borrowers would carry too much debt. Therefore, clients who were eligible for the Emergency Loan

and the Good loan then faced a tradeoff: they could take the Good Loan before the flood season

occurred and forgo the option of accessing additional liquidity in the event of a flood, or they could

preserve their credit access as a buffer against future flood risk. Clients who had access to the Good

8The danger level is not the water height at which the river overflows it banks, but at which there is estimated tobe high probability of significant property damage in the area.

9This second check was deemed necessary after piloting showed that for some branch service areas, a higher riverwater level was necessary to cause any risk of flood damage.

10Branch managers in the control group performed this same veto process for a future product rather than theemergency loan itself.

11This is also a standard practice for every other loan.12Of the 350,000 individuals in the data, approximately 165,000 (47%) were eligible for a Good Loan at some point

during the experiment. Of these, 66,000 (40%) were also eligible for the Emergency Loan.

7


Loan but not the Emergency Loan did not face this tradeoff. This creates an additional feature of

the experiment that I can exploit. Namely, I can compare these two groups to determine whether

households choose to preserve credit access as a buffer stock against risk. Figure 1 summarizes

borrower choices related to the Good Loan and Emergency Loan.

3 Theory

Framework For Effect of Guaranteed Credit

Clients are informed about their eligibility for the Emergency Loan in April, before decisions need

to be made on inputs for the coming Aman season (e.g. land to cultivate, inputs to use, business

investments), and how much to borrow to finance these choices. After making these decisions, the

cropping season commences and flooding either does or does not occur. If flooding does occur, each

eligible borrower is informed that the Emergency Loan is available for them to access. During this

period, borrowers make decisions on whether or not to take the Emergency Loan (if it is available),

and whether or not to repay existing loans. Later, borrowers move into the dry season and choose

to repay any loans taken after flooding.

Therefore, it is useful to categorize client decisions into three periods: choices made after being

informed about their Emergency Loan eligibility but before the realization of any flooding (first

period decisions), choices made after any flooding has occurred (second period decisions), and

choices made in the dry season (third period decisions).

First Period Decisions

1. Productive Investments: Households decide how much to invest in production, whether in

agricultural land and inputs, or in other business investment.

2. Dabi Loan Uptake: Each member will decide whether, and how much they wish to borrow

before the start of the Aman season.

3. Good Loan Uptake: For members who are eligible to take a Good Loan, they will decide

whether or not to take this additional credit to invest for the Aman season.

Second Period Decisions

1. Emergency Loan Uptake: In the event of a flood, borrowers will make the decision about

whether to take an Emergency Loan.

2. First Period Loan Repayment: Once borrowers choose whether or not to take the Emergency

Loan, they will need to decide how (or whether) to repay the loans they have.

8


Third Period Decisions

1. Second Period Repayment: Borrowers choose whether to repay the Emergency Loan if they

took one in the second period. This decision will also depend on whether or not the household

defaulted in the second period.

Below, I present a simple model that seeks to provide a framework for understanding how the

extension of a guaranteed credit could impact each of these decisions in turn.

Baseline Model

The model13 has three periods that correspond to a planting, harvest, and post-harvest periods, and

incorporates risky production, a credit market with constraints, and assumes that no insurance is

available. For ease, I limit the harvest realization to two possible states, s ∈ {G,B} that are realized

in time period two and occur with probability πs (later defined as πB = q and πG = (1−q)). Further,

I assume that the only source of credit available to a household comes from the MFI. Preferences

are over consumption (c) with discount factor β:

u(c1) + β∑

s∈G,B

πsu(c2s) + β2∑

s∈G,B

πsu(c3s)

A household starts with exogenous cash on hand Y and also has access to a risk free asset b1

which it can buy (up to a limit) or sell on the market at interest rate R (therefore positive values

of b represent net borrowing while negative values represent net saving). The household also has

access to a concave production function msf(x), which takes input x and provides output in the

second period. The production function has a state dependent marginal product ms which changes

with the realized state s. In period two, the state of the world is resolved and the household

decides whether to repay its initial loan with interest (Rb1) or default by paying zero. I also allow

for borrowing in the bad state of the world b2B, which is made available with the introduction of

the Emergency Loan (to simply the problem, I do not allow savings from period two to three,

but this assumption does not change the core results). In period three, the household pays (or

receives) return R on any period two loans, provided they have not already defaulted, and also

receive an exogenous risk free income (I). Finally, households that default are penalized K, which

is the household specific loss in utility from losing access to future dealings with the MFI. The basic

household problem can then be stated as:

maxx,b1,b2B ,D,ND

{u(c1)+∑

s∈G,B

max{βπsu(c2s|ND) + β2πsu(c3s|ND),

βπsu(c2s|D) + β2πsu(c3s|D)−K}} s.t.

13Based on model from Karlan and Udry (2015)

9


c1 = Y − x+ b1

c2G = 1 [ND][mGf(x)−Rb1

]+ 1 [D] [mGf(x)]

c2B = 1 [ND][mBf(x)−Rb1 + b2B

]+ 1 [D]

[mBf(x) + b2B

]c3G = I

c3B = 1 [ND][−Rb2B + I

]+ 1 [D] [ I ]

x ≥ 0

b1 ≤ B1 , (λ1)

b2B ≤ B2 , (λ2)

Where D and ND stand for default and no default respectively, cts and bts are consumption and

borrowing choice in the corresponding time period and state, x is inputs, Y is exogenous first period

wealth, and I is the exogenous third period income.

A household can borrow up to Bj in each period where borrowing is possible (where if B is equal

to zero there would be no access to credit). To begin, I will assume B2 = 0, meaning there is no

credit available in the bad state. I also make a few additional simplifying assumptions. First, I

assume that it is never optimal for a household to default on their loan when the good state is

realized (s = G). This assumption rules out households that always default and therefore took first

period loans in bad faith. Second, I normalize the marginal product of x is zero in the bad state,

i.e. mB = 014.

The rest of this section is organized as follows. First, I start by separately describing the

optimal borrowing and input choices assuming households do not default and then again assuming

households will default in the event of a shock. Second, I compare these two scenarios and find

the condition that will lead to the household to choose to repay or to default. Third, I allow for

bad state borrowing and observe how the relaxation of this constraint changes household choices

of inputs, borrowing, and the choice to default. Finally, given the expected effect on households’

decisions, I examine the implications of extending bad state borrowing on the performance of the

lending MFI.

14Note that this normalization also implies a shift in the utility function such that the utility of a negative valuedoes not imply zero or negative utility.

10


No Default

In this section, I derive the optimal choice of first period input use and borrowing assuming that

the borrowing will not default in the event of a shock. The household’s problem is:

maxx,b1

u(Y − x+ b1) + qβu(−Rb1

)+ (1− q)βu

(mGf(x)−Rb1

)+

qβ2u(I) + (1− q)β2u(I) + λ1[B1 − b1](1)

Where λ1 is the Lagrange multiplier associated with the first period borrowing constraint. Opti-

mizing equation 1 implies that the input x is purchased until the following condition is satisfied:

mG∂f

∂x= R

[q

1− qu′(c2B)

u′(c2G)+ 1

]+

λ1β(1− q)u′(c2G)

(2)

Under a scenario without risky production or credit constraints, the agent would invest in x

until the marginal product equaled the return on the risk-free asset R. The equation above shows

us that there are two potential sources of distortion away from that standard result. The first

term in brackets above will be greater than one and reflects the presences of the risky production

technology that has a zero marginal product in the event of the bad outcome. Second, the first

period credit constraint could bind in which case λ1 > 0, which will also drive a wedge between the

marginal product of the input and R. Therefore, both potential distortions will lower the choice of

x relative to the unconstrained optimum.

Now, I move to examine the borrowing choice in period one. The first order condition implies

that the first period borrowing is chosen such that:

u′(c1) = βR[qu′(c2B) + (1− q)u′(c2G)

]+ λ1 (3)

Again, there are two potential distortions away from the optimum without risky production or

credit constraints. First, the gap between second period consumption in the bad and good state

(qu(c2B) and (1 − q)u(c2G)) will increase the size of the second term (due to concavity), and imply

reduced consumption in period one relative to a choice without risky production, which combined

with the reduction in inputs purchased, implies an overall reduction in borrowing as well. Second,

as before, if the first period borrowing constraint binds, λ1 will be positive and will also imply a

reduction in borrowing relative to the unconstrained optimum.

11


Default

In this section, I assume that the household will choose not to repay their period one loans if the

bad state occurs in the second period. Under this assumption, the household problem changes to:

maxx,b1

u(Y − x+ b1) + qβu (0) + (1− q)βu(mGf(x)−Rb1

)+

qβ2u(I) + (1− q)β2u(I) + λ1[B1 − b1] + λ1[0− b2B](4)

The fact that the household knows they will not repay their loans in the event of a shock,

changes the optimal use of inputs and borrowing in the first period. First, I can see that the

optimal choice of inputs is defined by

mG∂fG∂x

= R+λ1

β(1− q)u′(c2G)(5)

This condition implies that households that know they will default in the bad state of the world

will use inputs until the marginal return is equal to the interest rate R. The only distortion comes

from the borrowing constraint in period one (λ1). Similarly, borrowing will be chosen such that

the only consideration is equalizing marginal utility in period one with discounted marginal utility

in period two:

u′(c1) = (1− q)βRu′(cG2 ) + λ1 (6)

Repayment Decision

To examine the borrower’s repayment decision, I compare the utility for the household when they

choose to repay to the utility they receive under default. If a household chooses to repay, their

utility under repayment must be higher than their utility under default:

U repay ≥ Udefault

Which is given by:

u(c1r) + qβu(−Rb1r

)+ (1− q)βu

(mGf(xr)−Rb1r

)+

qβ2u(I) + (1− q)β2u(I)

≥

u(c1d) + qβu (0) + (1− q)βu(mGf(xd)−Rb1d

)+

qβ2u(I) + (1− q)β2u(I)− qK

(7)

12


Where an index of d or r signify the optimal value of the variable given repayment or default

respectively. To understand this decision, I examine the switch point where a household is indifferent

between repayment and default by setting these two expressions equal. In order to declutter the

expression, it is useful to define new terms. First I define M as the difference in utility between

default and repayment in the first period and in the second period under the good state, where

M > 0.15

Using this simplification and rearranging the initial condition I can define K∗:

K∗ =M

q+ β

[u(0)− u(−Rb1r)

](8)

Where K∗ is the cost of lost access to microfinance that would make a household indifferent

between repayment and default.16 If a household’s actual K is larger than K∗ they will repay and

if it is lower, they will default. Therefore, if I assume that K is a random variable defined by the

CDF FK , the proportion of households that will default after shock is given by FK(K∗).

Adding Liquidity in the Bad State

I now explore how the optimal choices of x and b1 change when the option to borrow in the bad

state in period two is added. Starting with the no-default case, the household’s problem is now

expanded to include the choice b2B:

maxx,b1,b2B

u(Y − x+ b1) + qβu(−Rb1 + b2B

)+ (1− q)βu

(mGf(x)−Rb1

)+

qβ2u(I −Rb2B) + (1− q)β2u(I) + λ1[B1 − b1] + λ2[B2 − b2B]

(9)

In order to understand how the introduction of borrowing after a shock influences decisions,

I assume that the first period borrowing constraint does not bind (i.e. λ1 = 0), which allows

first period choices of x and b1 to adjust rather than being fixed at the constraint. Under this

assumption, the optimal choice of x is determined by:

mG∂fG∂x

= R

[q

1− qu′(c2B)

u′(c2G)+ 1

](10)

15

M =[u(c1d) − u(c1r)

]︸︷︷︸First Period

+[(1 − q)βu

(mGf(xd) −Rb1d

)− (1 − q)βu

(mGf(xr) −Rb1r

)]︸︷︷︸Second Period Good State

The difference in these terms is only due to the different optimal choices of x and b1 in the first period, rather thanthe repayment (or non-repayment) of loans. Therefore, because I know that xd > xr and b1d > b1r, the utility receivedwhen a client defaults is higher than the repayment utility. Therefore M > 0.

16Note that K∗ is monotonically increasing in b1, implying the more indebted a household the higher value of Knecessary to ensure repayment.

13


Allowing borrowing after a shock in the second period will increase consumption in this state (c2B)

relative to the constrained case. Thus, u′(c2B) decreases as does the ratiou′(c2B)

u′(c2G)in equation 10.

This implies the entire RHS of the equation falls and that therefore that optimal first period input

use will rise.17

I use a similar argument for first period borrowing, where the gap between u′(c2B) and u′(c2G)

is reduced in the equation 11 below, which causes the entire RHS of the equation to fall. This in

turn implies and increase in period one consumption (and therefore an increase in borrowing).

u′(c1) = βR[qu′(c2B) + (1− q)u′(c2G)

](11)

Last, I examine what factors determine the choice of b2B. Because there is no uncertainty moving

into the third period, the optimal choice of bad state borrowing is defined by the standard condition:

u′(c2B) = βRu′(c3B) + λ2

Households will be more likely to borrow in the bad state if they have a particularly low value of

c2B or have a high value of c3B. Therefore, I would expect more demand for the Emergency Loan

from households that are hit hardest by a flood shock and those that have high expected future

income I.

Therefore, the model gives four predictions that result from extending a credit line in the bad

state of the world:

1. Consumption increases after a shock

2. First period investment increases

3. First period borrowing increases

4. Probability of taking the Emergency Loan will be higher among households that experience

heavy damage from flooding or those with good post-Aman income opportunities

If I consider the case of households that default after a shock, it is easy to see that only prediction

1 will carry through. These households will indeed choose borrow in the bad state and therefore

increase their consumption as they do not plan to repay the loan. However, because they already

planned to default if a shock occurred, neither ex-ante input choice or first period borrowing will

be impacted by changes in the level of c2B. Further, I can see that the optimal bad state borrowing

amount will always be to take the maximum allowed, b2B = B2, as there is no cost of repayment

when already under default.

17Appendix A shows a more formal derivation of the comparative statics of x and b1 with respect to b2B .

14


Interaction with the Good Loan

I now consider the situation faced by clients who also have access to the Good Borrower loan.

Without the Emergency Loan, these households solve the same baseline model as above, with the

only difference being that their first period borrowing constraint is 1.5B1.18 However, with the

introduction of the pre-approved Emergency Loan, which is mutually exclusive with the Good

Loan, the problem facing these households changes. The borrowing constraints facing a household

in this situation are:

b1 ≤ 1.5B ,

b2B ≤ 0.5B

b1+b2B ≤ 1.5B ,

Now, any borrowing above B in the first period (i.e. using the Good Loan), comes at the

expense of available liquidity after a shock. It is for borrowers in this position that the problem

of credit line preservation becomes salient - households must now consider the value of preserve

their credit line for a time of need is, and whether or not this is worth forgoing current period

investment. The constrained maximization problem changes to:

maxx,b1,b2B

u(Y − x+ b1) + qβu(−Rb1 + b2B) + (1− q)βu(mGf(x)−Rb1)+

qβ2u(I −Rb2B) + (1− q)β2u(I) + λ1[1.5B − b1]+

λ2[0.5B − b2b ] + λ3[1.5B − b1 − b2B]

To simplify expressions, I assume that the Emergency Loan credit availability (0.5B) would be

enough so that the borrower will not be credit constrained in the bad state of the world if they

maintain their full credit line (i.e. λ2 = 0). Under this assumption, the ex-ante input choice

optimality is now determined by:

∂fG∂x

= R

[q

1− qu′(c2B)

u′(c2G)+ 1

]+

λ1β(1− q)u′(c2G)

+q

1− q

[u′(c2B)− βu′(c3B)

u′(c2G)

](12)

The first two terms of the equation are the same as we have seen previously in equation 2.

However, the last term is new and reflects that fact that with the combined constraint, any bor-

18Note that this result relies on the implicit model restriction that households cannot both borrowing and save inthe same period.

15


rowing in period one now limits the ability to smooth consumption in the future bad state by using

some of the household’s credit line. If this cross-period constraint binds (λ3 > 0), then u′(c2B) and

βu′(c3B) will not be equalized and the numerator in the last term will be positive. This has the

effect of increasing the value of the right hand side of the equation implying that the increase in

ex-ante inputs will be lower than for a Good Loan eligible client who did not have access to the

Emergency Loan.

Turning to the first period borrowing choice, the condition is now:19

u′(c1) = βR[qu′(c2B) + (1− q)u′(c2G)

]+ λ1 + qβ

[u′(c2B)− βu′(c3B)

](13)

Here, again, there is an additional term reflecting the potential gap between period two and three

consumption in the bad state. As before, if the combined borrowing constraint binds then the

third term will be positive and this will imply that the increase first period borrowing will be lower

relative to a Good Loan eligible client that does not have access to the Emergency Loan.

These results imply that when we consider the impact of the new product, we should see on

average a stronger effect of the Emergency Loan product on ex-ante input use and borrowing among

clients that are not eligible for the good borrower loan than among those who are. Additionally,

the emergency loan product will reduce the probability that eligible clients actually take the Good

Loan if they expect to be credit constrained in the bad state of the world.20 Therefore, we get two

further predictions:

5. The treatment effect on first period investment will be lower among Good Loan eligible

clients.

6. The offer of the Emergency Loan will reduce the probability that eligible clients take the

Good Loan.

Repayment Decision with Guaranteed Credit

The goal here is to understand how allowing second period borrowing in the bad state changes

borrowers’ loan repayment decisions. Recall equation 8 that defined the value K∗, which is the

benefit of future access to microfinance that would make a households indifferent between repayment

and default. With the introduction of the Emergency Loan, this expression expands to include the

option to borrow in the second period bad state and therefore also repay, or not, in the third period:

K∗ =M

q+ β

[u(b2B)− u(−Rb1r + b2B)

]+ β2

[u(I)− u(I −Rb2b)

](14)

19Again, assuming λ2 = 020Because in reality the uptake of the Good Loan is a binary choice, the effect of the Emergency Loan on Good

Loan uptake will be weakly negative.

16


To see how the repayment rates change with the introduction of the Emergency Loan, we need to

sign ∂K∗

∂b2Bwhen evaluated at b2B = 0.

∂K∗

∂b2B=

1

q

∂M

∂b2B︸︷︷︸−

+β

[u′(0)− u′(−Rb1r)

(1−R ∂b1r

∂b2B

)]︸︷︷︸

−

+β2Ru′(I)︸︷︷︸+

(15)

The first and second term above are negative, which capture improved good state outcomes and

the reduced cost of repayment respectively when the Emergency Loan is available. However, the

last term is positive and captures the added benefit of default when given more credit. Therefore,

the overall effect on repayment is ambiguous.

MFI Problem

I now move beyond the household and consider the implications of offering guaranteed credit after

a shock from the MFI’s perspective. I assume that the lender is maximizing interest revenue minus

the cost of defaults. For simplicity, I ignore the cost of capital and assume loans are either repaid

in full (earning the MFI b(R − 1) in interest), or lost completely costing the branch the full loan

amount b. When a shock occurs, F (K∗) gives the proportion of borrowers who will default on their

loan. As before, I assume that there is no default under the good state. The MFI’s expected profit

from lending to a particular household (defined by parameters Y and I) is therefore given by:

Π = q [(1− F (K∗)) (R− 1)b− F (K∗)b] + (1− q)(R− 1)b (16)

We are interested in whether it is profitable for the MFI to extend additional, guaranteed liquidity to

borrowers after a shock has occurred. To explore what happens to expected profits with this policy

change, we can simple explore how equation 16 changes when the amount borrowed b, is allowed to

move from b1 to (b1 + b2B).21 The MFI will want to offer the Emergency Loan if ΠE ≥ ΠNE , where

E and NE stand for Emergency Loan and No Emergency Loan respectively. This is given by

q[(1− F (K∗E))(R− 1)(b1E + b2B)− F (K∗E)(b1E + b2B)

]+ (1− q)(R− 1)b1E

≥q[(1− F (K∗NE))(R− 1)(b1NE)− F (K∗E)(b1E)

]+ (1− q)(R− 1)b1NE

(17)

Where K∗E , K∗NE and b1E , b1NE represent the indifference points for repayment and optimal first

period borrowing choice with and without the Emergency Loan respectively. Rearranging equation

21I assume households will take the Emergency Loan when offered, as otherwise the expected profits do not change

17


17, we can write that profits will increase if

q(R− 1)[(1− F (K∗E)(b1E + b2B)− (1− F (K∗NE)(b1NE)

]︸︷︷︸A

+

q[F (K∗NE)b1NE − F (K∗E)(b1E + b2B)

]︸︷︷︸B

+

(1− q)(R− 1)(b1E − b1NE)︸︷︷︸C

≥ 0

(18)

In equation 18, term A captures the change in profits from repayments. Because we know

that b1E is at least as large as b1NE , then b1E + b2B ≥ b1NE unambiguously.22 However, as we saw in

equation 15, the effect of the Emergency Loan on K∗ is ambiguous, therefore it is unclear whether

(1 − F (K∗E)) is greater or less than (1 − F (K∗NE)). If the offer of the Emergency Loan improves

repayment rates (∂K∗

∂b2B< 0) then A is clearly positive. However, the offer worsens repayment rate,

then the sign of A is ambiguous.

Similarly, term B captures the lost capital from defaults. We know that b1E + b2B ≥ b1NE , but

it is unclear whether F (K∗NE) is greater or less than F (K∗E). Therefore, as before, the sign of

term B depends on what the effect of the Emergency Loan is on repayment rate (i.e. the sign and

magnitude of ∂K∗

∂b2B). If ∂K∗

∂b2Bis positive, then this term is clearly negative and there will be larger

losses from default. However, is ∂K∗

∂b2Bis negative, then the overall sign of B is ambiguous.

Finally, term C captures profits when there is no shock. Again, this term is ambiguous. For

households without access to the Good Loan in the pre-period b1E ≥ b1NE . However, for households

with access to the Good Loan then b1E could be less then b1NE for clients who choose to preserve their

access to the Emergency Loan. The size of these effects and the number of households that are in

each situation will determine the overall sign of term C. Taking all three terms together, the overall

effect on MFI profits from offering the Emergency Loan is ambiguous, and will be determined by

i) the extent that the Emergency Loan changes households repayment rates positively and ii) how

the number of loans the MFI extends (including Dabi, Good, and Emergency Loans) changes as a

result.

4 Research Design

The impact of the Emergency Loan was tested using a randomized control trial with a sample of 200

BRAC branches. These 200 branches were randomly selected from a group of branches that satisfied

several criteria. First, I only included branches located in flood-prone areas. Second, I limited the

22This is clear for households without access to the Good Loan, however for households with access to the GoodLoan, the situation is less clear. Because the Good Loan and Emergency Loan are the same size by design, householdswith a preexisting Dabi loan will either be able to take a Good Loan or the Emergency Loan, leading to the sametotal borrowed amount. However, treated households may optimally increase their Dabi loan size (this is unlikely inthe first year of the program due to the timing of the pre-approval notification), in which case the borrowing amountwill again be larger.

18


sample to branches that were located within 15 kilometers of a river gauge run by the government’s

Flood Forecasting and Warning Center (FFWC) so that flooding could be monitored remotely.

Next, I analyzed 15 years of historical data from the FFWC river gauges and selected areas of the

country where flooding had exceeded the danger height levels at least twice. Last, I consulted the

Bangladeshi branch of the International Rice Research Institute (IRRI) and the BRAC branches

themselves to confirm that each branch’s service area had experienced flood damage in the past

six years. Figure 2 shows a map of the selected branches, their treatment status, and the matched

water level gauges. The selected branches are concentrated in four main regions, including the

Jamuna (Brahmaputra) basin, the Atrai river and Padma (Ganges) river basin, the Meghna river

basin, and the Feni river basin. A total of 100 branches were assigned to the treatment group, and

the remaining 100 branches were placed in the control group stratified by district. Table 3 provides

descriptive statistics from households sampled from the treatment and control branches and p-

values for the differences between these groups. The table shows that the randomized branches are

largely balanced on baseline observables.

The experiment began in April 2016 when the Emergency Loan eligibility lists were created

in each of the 200 experimental branches. Each branch manager could then review the lists and

remove up to 10% of the eligible borrowers based on their knowledge of borrowers’ behavior. The

final eligibility lists were then sent to BRAC headquarters for data keeping and to verify that no

more than 10% of borrowers had been removed from the original lists. Once finalized, referral

slips (see Figure 3) were created for each eligible borrower in the branch. Each slip contained

the borrower’s name, BRAC identification numbers, and details on the Emergency Loan including

the amount they had been pre-approved to borrow, the conditions when the loan would be made

available, and the fact that they would lose their eligibility status if they took a Good Loan. The

top half of the slip was kept by the borrowers to serve as “proof” of their eligibility status and

to serve as a reference about the details of the loan. The bottom half of the slip was filled out

with the borrower’s information and phone number to help the branch management contact eligible

borrowers after a flood.

The referral slips were distributed throughout the month of April during the normal VO meet-

ings for each branch. At the end of each meeting the loan officer distributed the referral slips to each

eligible borrower and read a script that explained the purpose and the key features of the product.

The concept of pre-approval was emphasized repeatedly because the idea was new within BRAC

microfinance operations. Borrowers were asked questions about the Emergency Loan to confirm

their understanding and time was given to answer any questions that eligible clients had about

the product. Random branch visits during June of 2016 confirmed relatively good execution of

pre-approval notification. Almost all borrowers had received the referral slips and understood that

the Emergency Loan was available in case of flooding. There was some heterogeneity in borrowers

understanding of the more nuanced details of the loan, including pre-approval and conflict with the

Good Loan. This was largely driven by difference in the quality of branch management.

During the Aman season, the FFWC flood water gauges were monitored every day for high water

19


levels. When a gauge showed water levels crossing the danger level, a BRAC research employee

(designated a “sector specialist”) was asked visit the branches matched with the gauge. They

mapped the area within each branch that had been affected by flooding based on conversations

with local officials. If the reported amount exceeded 20%, the branch was activated (importantly,

the sector specialists did not know about the 20% threshold needed to activate each branch). The

branch manager was instructed from headquarters to notify all eligible borrowers that Emergency

Loans were available. Borrowers were notified through their normally scheduled VO meetings or,

in cases where VO meetings were suspended because of flooding, by calling clients directly and

passing information through BRAC’s social network. Additionally, eligible clients were reminded

about the Emergency Loan’s availability at every subsequent VO meeting until the expiration of

the offer in November.

Over the course of the 2016 Aman season 92 branches were activated, 40 control and 51 treat-

ment.23 However, 2016 was not a major flooding year and the water levels in the majority of

activated branches did not cause widespread damage. As a result, BRAC decided to continue

piloting the Emergency Loan for a second year in 2017. From 2016 to 2017, the experimental

protocol remained the same. Only small improvements were made to the loan officer’s description

of the product. However, 14 branches (7 treatment, 7 control) were removed from the experiment

from 2016 to 2017 due to changes in the local topography (new dams and roads) that reduced

the probability of local flooding in these regions to almost zero. These 14 branches were replaced

with back-up branches that had been already pre-selected in the initial selection process described

above. The new branches were randomized into treatment and control in February 2017. In 2017,

136 branches were activated, 73 control and 63 treatment. Flooding in 2017 was in general more

severe than the previous year, and several locations suffered significant damage to crop land and

physical structures.

5 Data

The data used in this analysis comes from two primary sources. First, I use BRAC’s administrative

loans and savings records for all clients in the experimental branches. This dataset reveals every

client’s borrowing behavior, including decisions to take loans, loan repayments and savings activi-

ties. Detailed repayment and savings data are available from April 2016 until January 2018, while

loan disbursals data extends back for 1-7 years depending on the branch.24 Within the loans data

set, we observe approximately 350,000 unique individuals and 1.3 million unique loans. Eligibility

for the Good Loan, which was not included in this data set, was compiled separately by BRAC for

23The discrepancy in activation rates is likely due to random chance (the difference in means is not statisticallysignificant). Much effort was put in to ensure that control and treatment branches flood activation procedure wasfollowed in the same way, and this policy way reinforced when the difference in activation rates emerged early in2016.

24Certain BRAC branches began digitizing data earlier than others and 2) some branches in the experiment werefounded relatively recently.

20


the purposes of this research.25 Figure 4 shows a timeline of uptake of the three loan types studies

in this research (Dabi Loan, Good Loan, and Emergency Loan) over the periods for which data is

available, with the Aman growing season in 2016 and 2017 shaded in gray.

Second, I use survey data collected from 4,000 BRAC clients and 800 BRAC staff drawn from

the 200 experimental branches. For the borrower survey, three village organizations (VOs) were

randomly selected from each branch. Fifteen eligible borrowers and five ineligible borrowers were

randomly selected within each VO. Three rounds of data collection took place: a baseline survey

was conducted in April 2016 before borrowers in treatment branches were informed about their

eligibility status; a follow-up survey was implemented in December 2016 after the first flooding

season, and a second follow-up took place in December 2017 after the second flooding season.

Survey rates were very good, 99% at the first follow-up and 98.9% at the second follow-up.26

The household surveys focused on both agricultural and non-agriculture business investments and

outputs, consumption, asset holdings, and household perceptions of and response to any flooding

that occurred in the area. The surveys of BRAC’s administrative staff included 4 branch-level

managers (both in and outside of microfinance operations) and asked about their perceptions of

flood risk, the most important local income generating activities, and their perceptions of overall

local flood damage in the branch service area.

6 Results

Emergency Loan Take Up

I first examine the decision to take the Emergency Loan after a flood shock. In both years uptake

of the Emergency Loan among eligible households was quite low. In 2016, only 2.9% of households

chose to take the loan, likely because the floods that year were not particularly severe in most

locations. In 2017, floods were much more damaging and uptake of the Emergency Loan increased to

5.4%. It is important to note that these low take-up rates do not necessarily imply that households

did not value or benefit from the Emergency loan’s availability. While I address this point in

more detail below, households can respond to the offer of a loan before flooding has even occurred.

Indeed, the Emergency Loan stimulates higher investments and greater output, suggesting it offers

important protection in the pre-period against low probability shocks. Furthermore, low ex-post

uptake of this product is not entirely unexpected because flood damage is highly idiosyncratic

within a branch service area such that certain villages may be dramatically affected while others

villages within the same branch will not be hit at all.

Table 5 reports which household characteristics correlate with take-up among the set of house-

holds that were offered the Emergency Loan (i.e. those that were in a treatment branch after a

flood). Considering first baseline characteristics, column 1 shows that households that took the

25Due to uncertainty about whether the project would continue for a second year, this data is missing for fivemonths between November 2016 and March 2017 while the decision to extend the experiment was being made.

26Survey rates were helped tremendously by BRAC’s network which enabled easy tracking of households thatrelocated within and between communities.

21


Emergency Loan are observably quite similar to households that did not on most dimensions (risk

aversion, time preferences, flooding history and income). Column 2 explores correlations between

uptake and households status after a flood. I see higher take-up among households that were less

well prepared for a flood and among those that experienced higher levels of distress. Furthermore,

figure 5 highlights lower yields among households that took the Emergency Loan. Finally, figure

6 shows that there is no significant difference in the probability of Emergency Loan uptake by

borrower credit score. Overall, the results suggest that the more vulnerable and worst affected

households are more likely to take advantage of the guaranteed credit offer, results that are largely

consistent with the theory.

Estimation Strategy

To estimate the effects of guaranteed credit lines on household level outcomes, I will compare

eligible BRAC microfinance members across treatment and control branches. Eligible clients in

control branches are those who had credit scores high enough to qualify for the Emergency Loan

had they been in a treatment branch. The baseline specification for household outcomes is therefore:

Yibdt = treatmentibdβ + αd + φt + Xibdγ + εibdt

Where Yibdt is an observed outcome for an eligible household i in branch b and district d during

year t. I regress each outcome on an indicator for treatment, a district fixed effect (the stratification

variable), a year fixed effect, and a vector of baseline controls to increase precision.27 Data from

both years of the experiment are pooled together (unless noted otherwise) and standard errors are

always clustered at the branch level.28 For “ex-post” outcomes that occur after the flood season, I

run the same regression with an additional indicator for flooding during the growing season.

The same basic procedure is largely followed for MFI level outcomes (e.g. loan uptake decisions,

repayments) but with a few modifications. Because I examine observations at the branch-month

level, and thus add month m fixed effects in addition to year and district fixed effects.29

Ybdmt = treatmentibdβ + αd + φt + ρm + εbdmt

Credit Line Preservation

As mentioned above, low take-up does not necessarily reflect the value that households attribute

to the Emergency Loan. Gains from the loan can be reaped even if households decide not to take

the loan after the uncertainty about the flood is resolved. Access to the loan improves welfare by

reducing households exposure to the downside risks associated with severe flooding. To test whether

27Controls include land owned by the household, household size, and head of household age and education unlessspecified otherwise

28Appendix B accounts for possible differential selection into eligibility in 2017. Results are stable when excluding2017 data or when instrumenting for eligibility using branch treatment status.

29Some regressions have only a single observation per year, in which case month fixed effects are dropped. Notethat this dataset does not contain baseline controls and hence they are not included in the regression

22


households recognize this crucial feature of the Emergency loan, I investigate two phenomenon.

First, I document whether households choose to preserve their credit access to insure themselves

against bad times. Next, I investigate whether households invest more in the pre-period because

they know they have a recourse to an additional loan in the event of a flood.

To investigate credit preserving behavior, I take advantage of the tension between the Emergency

Loan and the Good Borrower Loan: households that take the Good Loan in the pre-period lose

access to the Emergency Loan. Eligible households have to choose whether to take a Good Loan

and forgo the Emergency Loan should a flood occur, or decline the Good Loan in order to preserve

the option to take the Emergency Loan after a shock. According to the theoretical model, forward

looking households will want to preserve credit access as a buffer against this risk. I test this

prediction by comparing the probability of taking a Good Loan among Good Loan eligible clients

in control branches compared to treatment branches in the pre-flood period. For this analysis I use

BRAC’s administrative data that captures loan disbursals and repayments.

Table 6 shows the results from the cross branch comparison of Good Loan eligible borrowers

(the regressions are run at the branch MFI level). Column 1 shows that the availability of the

Emergency Loan reduces the probability of taking a Good Loan by 2 percentage points, or 15%

in treatment branches. Column 2 and 3 examine the extent to which this effect changes based on

a measure of the need for liquidity, and by the perceived risk of local flooding. I proxy the need

for liquidity by areas that report farming as the primary occupation, where significant investments

are needed in the pre-period to prepare the seedbeds for cultivation. I do not see any significant

change in the treatment effect in areas where the primary occupation is farming. However, areas

that have a higher perceived flood risk are even less likely to take the good loan. Together these

results shows that a significant number of households that would have exhausted their available

credit absent the Emergency Loan, choose instead to preserve it. Furthermore, areas that face

higher risk are more likely to preserve their credit access, confirming that at least some households

view guaranteed credit access as offering effective insurance against shocks.

The fact that households are willing to give up investment in the pre-period suggests that the

value of preserving the guarantee is substantial for at least this subset of households. Those that

forgo the Good Loan in order to preserve access to the Emergency Loan are giving up certain credit

today in order maintain credit in the future that will only be made available with some probability.

If I assume that households were able to correctly predict the probability a loan would be offered

(54% over the two years of the study), this implies that households value access to credit in the

event of a flood at a minimum of 1.85 times the value of certain credit in the pre-period.

In order to understand which borrowers are more likely to actively preserve their credit access,

I estimate a local average treatment effect across bins of the Emergency Loan credit score. Figure

8 plots the treatment effect on Good Loan uptake by credit score bin for eligible clients. There

appears to be some evidence of heterogenous treatment effects: the reduction in the probability of

taking a good loan among the eligible population is highest among those with especially high credit

scores. Column 1 of Table 18 fits a linear trend to this relationship and shows that this effect is

23


(marginally) statistically significant. This suggests that clients with the best repayment histories

are more likely to preserve credit access to hedge against future shocks. We might expect this result

if clients with higher credit scores have lower discount rates or they are less present biased. Such

households would likely make more timely payments (hence the higher credit scores) and be willing

to preserve credit access.

Ex-Ante Household Investment

Recall from theory that the extension of a guaranteed credit line is designed to mitigate the adverse

consequences of a shock, thereby encouraging households to invest more in the pre-period. I focus

on changes to agricultural investments because it is the most important income generating activity

for the majority of rural households in Bangladesh. Moreover these investments are more likely to

be exposed to flood shocks and sensitive to interventions that reduce household flood risk. I also

investigate the impacts on non-agricultural business investments because the sample is drawn from

microfinance clients that are more likely to be business owners and less likely to own land than the

general rural population (48% of surveyed households planted crops during the 2015 Aman season).

I begin with Table 7, which shows the amount of land devoted to agriculture during the rainy

season. The first three columns separately identify the impact for three different types of land

tenure, namely owned, rented, and sharecropped land, while column 4 aggregates these three mea-

sures. The last column is a binary indicator for planting any crops during the season. Households

that knew they were eligible for the loan increased the amount of land they rented by 30%, and

overall cultivation by 15%. Neither owned nor sharecropped land showed any significant change.

This result is not altogether surprising because finding additional land to rent is relatively straight

forward. Conversely, expanding the cultivation of owned land would require farming previously

fallow land or purchasing additional crop land, which is more costly and requires more planning.

Similarly, sharecropping contracts become relatively less attractive because farmers ability to re-

duce their exposure to risk can now be fulfilled by the Emergency loan. Finally, along the extensive

margin, the number of eligible households planting crops increased by approximately 4 percentage

points. This represents a 10% increase in the probability that a household cultivates crops during

the Aman season.

With an expansion in cultivated land, total input use is likely to increase mechanically. However,

households might also increase the intensity of input usage in response to reduced exposure to risk.

The first four columns of Table 8 present the effects of the intervention on inputs applied to

cultivated farm land. Columns 1 and 2 show the amount of fertilizer and pesticides applied per

acre of land. While both variables have positive point estimates, neither are statistically significant.

Similarly, columns 3 and 4 show that the amount of money spent on seeds and all other inputs

per acre increase but remain insignificant. At a minimum, these results indicate that treatment

households are maintaining normal levels of input usage per acre despite the overall expansion

of cultivated land. Finally, column 5 of Table 8 examines the amount of investments in non-

agriculture business. We see a marginally significant increase of 29% ($11 USD) over the control

24


group.30 However, this last result is sensitive to the regression specification and only becomes

significant in the second year of the experiment.

These initial results are consistent with the theory that guaranteed credit lines can increase

investments by providing effective insurance against floods. However, to confirm that the product

is operating on farmers’ perceptions of risk, I investigate whether the treatment effects are higher

among the most risk averse households (as measured at baseline).31 These households represent

a meaningful share of the sample (27% of households exhibit the highest level of risk aversion),

and generally invest less at baseline. Table 9 and 10 report these results, where the measure for

risk aversion is normalized to a 0-1 scale (one representing the most risk averse households and

zero the most risk loving). From Table 9 we see that all the point estimates on the interaction

terms between risk aversion and treatment are positive. However, they are only significant for

rented and total land cultivated (which is where I documented the strongest impacts previously).

I also investigate the impact for risk averse households specifically by running a joint test on the

treatment and interaction terms. Here I find significance for the treatment effect on rented land in

addition to total land. Similarly, in Table 10 the interaction term is positive for fertilizer, pesticide,

and for non-agricultural investment. The joint tests indicate that pesticide application and non-

agricultural investment have increased significantly for most risk averse households. Overall these

results suggest that guaranteed credit lines are encouraging investments by reducing households

exposure to risk. The fact that risk averse households tend to undervest in general suggests this

product is particularly valuable at correcting a negative distortion for this subgroup.

The results on increased investment and credit preservation suggest that households perceive

the Emergency Loan as reducing their exposure to risk. However it is possible that households

choose not to take the Emergency Loan because they learn ex-post that it would never be useful to

them (in this season or in any future one). This might be driving the low take-up results we saw

above. If households are learning this, we would to expect see their 2017 Aman season investments

decrease to pre-treatment levels because their choice not to adopt in the future eliminates the risk

reduction benefit of guaranteed credit. To test this theory, I examine how investment decisions

change in the second year of the experiment based on whether households experienced a flood

shock in the first season. If flood afflicted households learn that the Emergency Loan does not

insure them against negative outcomes, these households should have a smaller treatment effect

on investment relative to treatment households that did not experience a flood shock. However, if

households still perceive the offer of guaranteed credit as reducing the downside risk of flooding,

then investment should be the same (or larger) when compared to households that were not flooded

in the first year.32

30Business investment was measured by the total value of newly purchased (or repaired) business assets.31Risk aversion was measured by asking borrowers a series of choices between a certain payout and a larger, but

uncertain payout. Each successive choice increased the probability that the uncertain payout would be realized (seeSprenger 2015 for more details). The resulting risk aversion spread was normalized to a zero to one scale so that themost risk averse households have a value of one and the most risk loving a value of zero.

32A possible confounding factor is that the extra credit afforded by access to the Emergency Loan in the first yearcould itself impact investment decisions in year two. However, this effect will have a minimal role due to the fact

25


Table 11 demonstrates how flooding in the first year affects different investment categories.

First, we can see that being flooded in the previous year does seem to have negative consequences

for control households’ investments in the current year. In particular, control households are ten

percentage points less likely to cultivate crops in areas that were flooded in 2016. However, the

treatment effect on investments does not appear to be different for treatment households that were

flooded in the first year relative to treatment households that were not. The interaction term is

generally small in magnitude and not statistically significant for any outcome. Overall, this suggests

that households that experienced flooding in 2016 still perceive the Emergency Loan as offering

viable protection against some flood risk.

Ex-Post Household Outcomes

Next I turn to examine how the Emergency Loan affects households after the Aman season, both

in areas that experience flooding and those that do not. Recall from the model that offering the

Emergency Loan will affect households differently depending on the state of the world. In the event

of a flood, the emergency loan becomes available and treatment households will have access to more

liquidity than control. If a flood does not a occur, increases in investment before the Aman season

will translate into improved output.

I examine the effect of treatment on four household outcomes: log weekly consumption per

capita, log income during the previous month, crop production from the Aman season, and the

number of livestock animals owned by the household.33 Table 12 shows the results of regressing

these outcomes on an indicator for treatment, an indicator for experiencing a flood shock during the

growing season, and an interaction between the two.34 The coefficient on treated captures the effects

on household outcomes from increases in ex-ante investment only. Absent a flood the only difference

in outcomes between treatment and control households will stem from changes in investment in the

pre-period. In contrast, the interaction term will capture the effect of both channels on household

outcomes. After a flood, treatment households will have access to any output the flood did not

destroy, and to the Emergency Loan should they choose to use it for recovery.

In branches that did not experience flooding, treated households display the same levels of

consumption, income, and livestock ownership as control households (Table 12). However, there is

a significant 28% increase in crop production which aligns with the pre-period results documenting

additional crops being sown by treatment households. This suggests that households reap the ben-

efits of greater investments absent a flood even though this does not translate into higher levels of

measured consumption or asset holdings. In branches that did experience a flood, treated house-

holds experience a rather large 10% increase in consumption compared to control households.35

However, their production is affected by the flood, losing almost 80% of the gains they reap when

that only 2.9% of eligible households took the Emergency Loan when offered in the first year.33The estimation for log consumption adds week interviewed fixed effects because of holidays that occurred over

the survey period which changed consumption patterns for some households.34See Appendix B for ex-post results pooling both flooded and non-flooded branches.35The p-value for the joint test is found in the bottom row of table 12

26


a flood does not occur (Column 3). These losses are proportionally much larger than those ob-

served in the control group, suggesting that treatment households expand cultivation on land that

is particularly susceptible to floods. Finally, we see a large increase in the number of livestock

among treatment households (Column 4). This suggests that the availability of the Emergency

Loan allows households to maintain their asset levels after an income shock.

There is a concern that multiple shocks may reduce the usefulness of credit as a risk mitigation

tool if households use their entire available credit line, thereby eliminating the products consump-

tion smoothing benefits. Table 13 examines this hypothesis. To do so, I expand the regression

specification from Table 12 to include an indicator for wether the household experienced flooding

in both years, and an interaction of this indicator with treatment. The experiment was only con-

ducted over two years, so multiple shocks can only be picked up for households that experience

flooding in both 2016 and 2017. The results confirm that experiencing successive shocks reduces

food consumption by 19%, but has little observable impact on the other three outcomes. This

suggests that multiple shocks are indeed harmful to households well being, even if the channels

through which this occurs are unclear. Next, to determine whether the usefulness of guaranteed

credit is reduced after successive shocks, I examine the interaction of the double flood indicator and

the treatment indicator. These coefficients are all statistically insignificant, but a joint test of all

the treatment coefficients shows that treatment households are still better off after a double shock.

Overall, this suggests that the gains in consumption and asset preservation due to treatment are

not completely eliminated by successive shocks. However, it is worth interpreting these results with

some caution because the 2016 shock was not particularly damaging, and may not reflect responses

to larger shocks.

Finally, Table 14 explores how the availability of the Emergency Loan changes outcomes for

households that earn income through the labor market rather than (or in addition to) agriculture

or a small business. The first column shows that laborers’ daily wage does not change in flooded

or non-flooded areas. In non-flooded areas however, the number of days employed elsewhere as a

day laborer decreases by 10%, likely a result of more households spending time on their own fields.

In flooded areas, we see a 30% reduction in day labor, which is substantially reversed in treatment

households (by approximately 20%). This result could be consistent with the fact that households

often use the Emergency Loan to re-plant the fields destroyed by floods, increasing the number of

opportunities for other laborers to find work. However, given the low rates of Emergency Loan

take-up, this channel is unlikely to explain the entirety of this effect.

Impact on MFI Operations

I conclude the analysis by investigating how BRAC branches perform when the Emergency Loan

is made available. As discussed in the theory section, it is unclear a-priori whether extending

guaranteed credit after a shock will help or harm overall branch performance. There are two key

outcomes that determine branch profitability: the number of loans disbursed and the repayment

rates of those loans. Therefore, to understand the effect of the Emergency Loan product I will

27


examine each of these outcomes in turn (recall we have already seen that the Emergency Loan

reduces the number of Good Loans disbursed).

I begin by examining whether offering the Emergency Loan increases the likelihood that bor-

rowers take an initial loan from BRAC. To do so, I examine the probability that a normal dabi

loan is taken in the pre-flood period among all members of the branch.36 The results in Table 15

show that treatment causes the probability of taking a dabi loan to increase by 11% (0.7 percentage

points) in the pre-period. However, it is possible that the increase in loan disbursement during the

pre-period came at the expense of future loans (for example, if households simply move up their

previously planned investment timeline). Figure 9 examines whether this is occurring by plotting

the monthly probability of dabi loan up-take by treatment status from 2015 until the end of the

study period. We can see that the probability of taking a new dabi loan is higher in the treatment

branches during the pre-period, but is otherwise fairly similar. This suggests that the extra dabi

loans disbursed in the pre-period represent additional loans that would not otherwise have been

disbursed. Finally, as with the Good Loan analysis, I examine whether the increase in Dabi Loan

uptake differs across credit scores. Figure 10 and column 2 in Table 18 shows that the increase in

Dabi Loans (unlike the reduction in Good Loan uptake) does not differ by credit score.

In addition to loan disbursals, impacts on repayment rates are critical to establish the sustain-

ability of the Emergency Loan. Table 16 shows how the probability of a missed payment differs

between treatment and control branches in the pre-period and after a flood. The coefficient on

treatment shows that access to the Emergency Loan has no effect on repayment rates in the ab-

sence of a shock. Looking at the coefficient on flooding, we see that flooding increases the number

of missed payments by approximately 3.9 percentage points (40% percent) in control branches.

However, in treatment branches this effect is overcome by a reduction in missed payments of 4 per-

centage points, thereby returning repayment rates to approximately normal rates. Furthermore,

the repayment rate of the Emergency Loan itself is almost identical to other loans during the same

period (10% missed payments for the Emergency Loan as compared with 9.6% on all loans). This

result is even more meaningful when we remember that households that took the Emergency Loan

experienced greater damages from the flood. Overall, these results demonstrate that the availability

of the Emergency Loan improved repayment for the MFI in the aftermath of the flood (on a branch

wide basis).

Next, I look for heterogeneity in repayments rates by borrower credit score. Figure 11 and

12 illustrate how repayment rates differ by credit score and by treatment status. First, Figure

11 shows that the treatment effect on repayment rates37 is largest among clients with scores that

are close to the eligibility threshold of 77. The effect falls quickly at higher credit scores (column

3 of Table 18 shows that this heterogeneity is statistically significant). This decrease is likely

explained by the fact that borrowers with high credit scores already repay at such high rates that

further improvements are difficult. In Figure 12, we see that approximately 6% of payments are

36All members were included in the analysis so that the denominator of eligible borrowers remained constantthroughout the study time period and did not change in response to endogenous loan take-up decision.

37The estimated treatment effect is from regressions pooling both flooded and non-flooded branches.

28


missed among those with high credit scores, which is low enough that it may be difficult to improve

repayment rates significantly.

Overall branch profitability is derived from the number of loans disbursed and the repayment

rates on those loans. So far, we have seen that the effect on total loans disbursed is ambiguous

– a decrease in the number of Good Loans taken, but an increase in the number of regular dabi

Loans and new Emergency Loans – while the effect on repayment rates appears to be positive.

To capture the overall effect on the branch, we can directly compare the profitability of branches

that offered the Emergency Loan to those that did not. Table 17 shows the estimated effects of

treatment on three measures of MFI profitability: the net present value of each loan disbursed, the

monthly profitability of the branch in aggregate, and the per-member monthly profitability of each

branch.38 The first two results show positive point estimates, but neither is statistically significant.

However, column 3 shows a 4% increase in the per-person profits in treatment branches. In sum,

these results suggest a modest increase in branch profitability and allow us to say that branch

profitability was likely not harmed.

We can examine the extent to which the effects on profitability vary by borrower credit score.

Figure 13 plots the treatment effect on per-person profitability by credit score decile. We see that the

treatment effect is highest for clients with credit scores closer to the eligibility cutoff and decreases

steadily until it is negative for those with higher credit scores (column 4 of Table 18 show that this

heterogeneity is statistically significant). These results are consistent with the effects I have shown

previously. Clients with scores near the cutoff both have the highest improvements on repayment

rates and the lowest reductions in the probability of taking a Good Loan. In contrast, high credit

score clients make only modest improvements to their repayment rates while experiencing the

largest reductions in the probability of taking a Good Loan.

These results have interesting implications for the targeting of the Emergency Loan. The

Emergency Loan was targeted at the top 40% of borrowers based on a credit score reflecting their

past loan behavior. This system was designed to reduce the downside risk for the MFI in case

repayment rates from the Emergency Loan were low. However, the results suggest that BRAC

could do even better by lowering the eligibility threshold. Assuming the measured treatment

effects are continuous across the threshold, this would extend access to clients who are most likely

to improve MFI profitability. In contrast, restricting access to the Emergency Loan to clients with

the highest credit scores could lead to an overall reduction in branch profitability because they are

less likely to take the Good Loan, and their repayment rates do not have room to improve.

As a final check on MFI performance we can look at saving rates. BRAC benefits directly

from the amount of savings stored by clients at the branch. Table 19 shows how the savings rates

differ between treatment and control branches and their differential response to flooding. Column

1 shows that in the pre-period, where we might have expected a draw down in liquid assets, savings

rates do not differ between the two branches. However, column 2 shows that in the aftermath of

38To calculate net present value for each loan, I assume an annual cost of capital of 6%. Branch profit is calculatedas the sum of discounted repayments minus the cost of new disbursements, while per-member profitability takes thismeasure and divides it by the number of branch members.

29


a flood, eligible households are able to maintain higher savings rates by 45 taka on average (which

represents a 62% increase on the average transaction amount, but less than a 1% increase on total

savings). Column 3 shows that this effect does not vary by the level of localized damage inflicted

by the flood39.

7 Conclusion

Millions of households across the world are exposed to severe income risk. In many cases, these

households live in areas where insurance markets are non-existent and they have to resort to costly

coping mechanisms in order to survive. Under these circumstances, it becomes important to develop

tools that can decrease households’ exposure to risk and help them self-insure. I build on recent

literature, which suggests that uninsured risk is an important constraint and that existing insurance

products are inadequate, by offering a guaranteed credit line as a potential solution. I run a large

scale RCT to investigate whether guaranteed credit can successfully function as insurance in rural

regions of Bangladesh where annual flood risk is high. First, I show that households value this

product: when given the choice, many households choose to preserve their access to guaranteed

credit at the expense of additional liquidity in the pre-period. This behavior is consistent with

a model where households utilize their credit access as a buffer against the risk of future shocks.

Household that were informed about their guaranteed credit access also increased their investments

in productive (but risky) activities in the pre-period. These effects were concentrated among more

risk-averse households. This increase in investments translated into more production absent a flood,

and higher consumption and asset levels when a shock did occur.

I also show that the extension of a guaranteed credit line after a shock has modest, but largely

positive effects for MFIs. More members take loans in the pre-period in response to the added

security, repayment rates after a shock are improved, and savings rates increase. Therefore, at

least in this context, a product like the Emergency Loan slightly improves branch performance in

addition to benefiting clients. Provided that loan repayment rates remain similar in other settings,

this suggests that guaranteed credit can be offered by MFIs and does not require any third party

subsidies. This is appealing because MFIs are ubiquitous in low income countries and can easily

offer the product to many households using their existing infrastructure.

One question raised by the results is why a product like the Emergency Loan that seems to

benefit both households and the MFI has not already been adopted by the microfinance industry. I

suggest two obstacles that may prevent adoption. First, many MFIs do not keep adequate records

and lack the lending history necessary to create a credit score to target responsible borrowers.

The results are unlikely to generalize to lower performing clients and it is important to be able to

identify who these households are. Second, a guaranteed credit product does not necessary align

with the incentives facing branch level officials. Branch managers are commonly incentivized to

disburse a certain number of loans and to maximize repayment rates. However, in the aftermath

39Flood damage at the branch level was only collected in 2017, therefore column 3 only uses data from this year.

30


of an aggregate shock, a branch manager may be concerned that households are going to miss

payments on their existing loans, and a product like the Emergency loan will compound these

losses. This would increase the downside risk facing the branch, and could potentially jeopardize

the manager’s job. Therefore, there is likely be resistance from branch level staff to adopt similar

guaranteed products.

From a policy perspective, this research suggests that credit can be a useful tool to address

uninsured risk in places where traditional insurance markets have failed. With the growing fre-

quency and severity of weather shock due to climate change, adding an easily accessible tool that

helps households reduce exposure to risk is important. The tool I explore here (guaranteed credit)

is appealing because it is already understood in these environments and it widely used worldwide.

While the household impacts are similar to those documented in the index-insurance literature,

pre-approved credit has the advantage that it can be extended without requiring any commitment

from the beneficiary, bypassing many of the drivers of low demand for insurance. Moreover, because

the decision to utilize the additional credit is made after any shock damages have been realized,

households can optimally opt-in based on the ex-post costs and benefits. Therefore, guaranteed

credit can crowd-in ex-ante investment even if households choose not to use the product in the

aftermath of a flood.

References

Angelucci, Manuela, Dean Karlan, and Jonathan Zinman. 2013. “Win Some Lose Some? Evi-dence from a Randomized Microcredit Program Placement Experiment by Compartamos Banco.”Working Paper 19119, National Bureau of Economic Research.

———. 2015. “Microcredit Impacts: Evidence from a Randomized Microcredit Program Place-ment Experiment by Compartamos Banco.” American Economic Journal: Applied Economics7 (1):151–182.

Banerjee, Abhijit, Esther Duflo, Rachel Glennerster, and Cynthia Kinnan. 2015. “The Miracle ofMicrofinance? Evidence from a Randomized Evaluation.” American Economic Journal: AppliedEconomics 7 (1):22–53.

Banerjee, Abhijit, Esther Duflo, and Richard Hornbeck. 2014. “Bundling Health Insurance andMicrofinance in India: There Cannot Be Adverse Selection If There Is No Demand.” AmericanEconomic Review 104 (5):291–297.

Banerjee, Abhijit, Dean Karlan, and Jonathan Zinman. 2015. “Six Randomized Evaluations ofMicrocredit: Introduction and Further Steps.” American Economic Journal: Applied Economics7 (1):1–21.

Barboni, Giorgia. 2017. “Repayment Flexibility in Microfinance Contracts: Theory and Exper-imental Evidence on Take up and Selection.” Journal of Economic Behavior & Organization142:425–450.

31


Beaman, Lori, Dean Karlan, Bram Thuysbaert, and Christopher Udry. 2014. “Self-Selection intoCredit Markets: Evidence from Agriculture in Mali.” Working Paper 20387, National Bureau ofEconomic Research.

Brammer, H. 1990. “Floods in Bangladesh: Geographical Background to the 1987 and 1988 Floods.”The Geographical Journal 156 (1):12–22.

Breza, Emily and Cynthia Kinnan. 2018. “Measuring the Equilibrium Impacts of Credit: Evidencefrom the Indian Microfinance Crisis.” Working Paper 24329, National Bureau of EconomicResearch.

Bryan, Gharad, Shyamal Chowdhury, and Ahmed Mushfiq Mobarak. 2014. “Underinvestmentin a Profitable Technology: The Case of Seasonal Migration in Bangladesh.” Econometrica82 (5):1671–1748.

Burgess, Robin and Rohini Pande. 2005. “Do Rural Banks Matter? Evidence from the IndianSocial Banking Experiment.” The American Economic Review 95 (3):780–795.

Carroll, Christopher D. 1997. “Buffer-Stock Saving and the Life Cycle/Permanent Income Hypoth-esis.” The Quarterly Journal of Economics 112 (1):1–55.

Casaburi, Lorenzo and Jack Willis. 2018. “Time vs. State in Insurance: Experimental Evidencefrom Contract Farming in Kenya.” American Economic Review .

Clarke, Daniel J. 2016. “A Theory of Rational Demand for Index Insurance.” American EconomicJournal: Microeconomics 8 (1):283–306.

Cole, Shawn, Xavier Gine, Jeremy Tobacman, Petia Topalova, Robert Townsend, and James Vick-ery. 2013. “Barriers to Household Risk Management: Evidence from India.” American EconomicJournal: Applied Economics 5 (1):104–135.

Cole, Shawn, Xavier Gine, and James Vickery. 2017. “How Does Risk Management InfluenceProduction Decisions? Evidence from a Field Experiment.” The Review of Financial Studies30 (6):1935–1970.

Cole, Shawn A. and Wentao Xiong. 2017. “Agricultural Insurance and Economic Development.”Annual Review of Economics 9 (1):235–262.

Crepon, Bruno, Florencia Devoto, Esther Duflo, and William Pariente. 2015. “Estimating theImpact of Microcredit on Those Who Take It Up: Evidence from a Randomized Experiment inMorocco.” American Economic Journal: Applied Economics 7 (1):123–150.

Daley-Harris, Sam. 2006. “State of the Microcredit Summit Campaign Report 2006 — MicrocreditSummit Campaign.” Tech. rep.

de Janvry, Alain, Craig McIntosh, and Elisabeth Sadoulet. 2010. “The Supply- and Demand-SideImpacts of Credit Market Information.” Journal of Development Economics 93 (2):173–188.

Deaton, Angus. 1991. “Saving and Liquidity Constraints.” Econometrica 59 (5):1221–1248.

———. 1992. “Household Saving in LDCs: Credit Markets, Insurance and Welfare.” ScandinavianJournal of Economics 94 (2):253–73.

32


Dehejia, Rajeev, Heather Montgomery, and Jonathan Morduch. 2012. “Do Interest Rates Matter?Credit Demand in the Dhaka Slums.” Journal of Development Economics 97 (2):437–449.

Demont, Timothee. 2014. “Microcredit as Insurance: Evidence from Indian Self-Help Groups.”Tech. Rep. 1410, University of Namur, Department of Economics.

Dercon, Stefan. 2002. “Income Risk, Coping Strategies, and Safety Nets.” The World Bank ResearchObserver 17 (2).

Donovan, Kevin. 2016. “Agricultural Risk, Intermediate Inputs, and Cross-Country ProductivityDifferences.” Unpublished Paper, University of Notre Dame .

Emerick, Kyle, Alain de Janvry, Elisabeth Sadoulet, and Manzoor H. Dar. 2016. “TechnologicalInnovations, Downside Risk, and the Modernization of Agriculture.” American Economic Review106 (6):1537–1561.

Fafchamps, Marcel, Jan Willem Gunning, and Remco Oostendorp. 2001. “Inventories and Risk inAfrican Manufacturing.” The Economic Journal 110 (466):861–893.

Field, Erica and Rohini Pande. 2010. “Repayment Frequency and Default in Microfinance: Evidencefrom India.” Journal of the European Economic Association 6 (2-3):501–509.

Field, Erica, Rohini Pande, John Papp, and Natalia Rigol. 2013. “Does the Classic Microfi-nance Model Discourage Entrepreneurship among the Poor? Experimental Evidence from India.”American Economic Review 103 (6):2196–2226.

Fischer, Greg. 2013. “Contract Structure, Risk-Sharing, and Investment Choice.” Econometrica81 (3):883–939.

Gine, Xavier and Dean S. Karlan. 2014. “Group versus Individual Liability: Short and Long TermEvidence from Philippine Microcredit Lending Groups.” Journal of Development Economics107:65–83.

Gine, Xavier and Dean Yang. 2009. “Insurance, Credit, and Technology Adoption: Field Experi-mental Evidencefrom Malawi.” Journal of Development Economics 89 (1):1–11.

Gine, Xavier, X. 2009. “Innovations Insuring the Poor: Experience with Weather Index-BasedInsurance in India and Malawi — FARMD: Forum for Agricultural Risk Management in Devel-opment.” Tech. rep., International Food Policy Research Institute.

Gross, David B. and Nicholas S. Souleles. 2002. “Do Liquidity Constraints and Interest RatesMatter for Consumer Behavior? Evidence from Credit Card Data.” The Quarterly Journal ofEconomics 117 (1):149–185.

Guiteras, Raymond, Amir Jina, and A. Mushfiq Mobarak. 2015. “Satellites, Self-Reports, andSubmersion: Exposure to Floods in Bangladesh.” American Economic Review 105 (5):232–236.

Hazell, Peter B. R. 1992. “The Appropriate Role of Agricultural Insurance in Developing Coun-tries.” Journal of International Development 4 (6):567–581.

Hoddinott, John. 2006. “Shocks and Their Consequences Across and Within Households in RuralZimbabwe.” The Journal of Development Studies 42:301–321.

33


Janzen, Sarah A and Michael R Carter. 2018. “After the Drought: The Impact of Microinsuranceon Consumption Smoothing and Asset Protection.” Working Paper 19702, National Bureau ofEconomic Research.

Jensen, Nathaniel and Christopher Barrett. 2017. “Agricultural Index Insurance for Development.”Applied Economic Perspectives and Policy 39 (2):199–219.

Kaboski, Joseph P. and Robert M. Townsend. 2011. “A Structural Evaluation of a Large-ScaleQuasi-Experimental Microfinance Initiative.” Econometrica 79 (5):1357–1406.

Karlan, Dean. 2014. “Innovation, Inclusion and Trust: The Role of Non-Profit Organizations inMicrofinance.” Tech. rep., Innovations for Poverty Action.

Karlan, Dean and Nathanael Goldberg. 2011. “Microfinance Evaluation Strategies: Notes onMethodology and Findings.” World Scientific Book Chapters, World Scientific Publishing Co.Pte. Ltd.

Karlan, Dean, Jake Kendall, Rebecca Mann, Rohini Pande, Tavneet Suri, and Jonathan Zinman.2016. “Research and Impacts of Digital Financial Services.” Working Paper 22633, NationalBureau of Economic Research.

Karlan, Dean, Robert Osei, Isaac Osei-Akoto, and Christopher Udry. 2014. “Agricultural Decisionsafter Relaxing Credit and Risk Constraints.” The Quarterly Journal of Economics 129 (2):597–652.

Karlan, Dean and Jonathan Zinman. 2011. “Microcredit in Theory and Practice: Using RandomizedCredit Scoring for Impact Evaluation.” Science 332 (6035):1278–1284.

Khandker, Shahidur R., M. Abdul Khaleque, and Hussain A. Samad. 2011. Can Social Safety NetsAlleviate Seasonal Deprivation? Evidence from Northwest Bangladesh. Policy Research WorkingPapers. The World Bank.

Knutson, Thomas R., Joseph J. Sirutis, Gabriel A. Vecchi, Stephen Garner, Ming Zhao, Hyeong-Seog Kim, Morris Bender, Robert E. Tuleya, Isaac M. Held, and Gabriele Villarini. 2013. “Dy-namical Downscaling Projections of Twenty-First-Century Atlantic Hurricane Activity: CMIP3and CMIP5 Model-Based Scenarios.” Journal of Climate 26 (17):6591–6617.

Maccini, Sharon and Dean Yang. 2009. “Under the Weather: Health, Schooling, and EconomicConsequences of Early-Life Rainfall.” American Economic Review 99 (3):1006–1026.

Maes, J. and L. Reed. 2011. “State of the Microcredit Summit Campaign Report 2012.” Paper,Microcredit Summit Campaign.

McIntosh, Craig, Alexander Sarris, and Fotis Papadopoulos. 2013. “Productivity, Credit, Risk, andthe Demand for Weather Index Insurance in Smallholder Agriculture in Ethiopia.” AgriculturalEconomics 44 (4-5):399–417.

Mel, Suresh De, David McKenzie, and Christopher Woodruff. 2011. “Enterprise Recovery FollowingNatural Disasters.” The Economic Journal 122 (559):64–91.

Miranda, Mario J. and Katie Farrin. 2012. “Index Insurance for Developing Countries.” AppliedEconomic Perspectives and Policy 34 (3):391–427.

34


Mobarak, Ahmed Mushfiq and Mark Rosenzweig. 2012. “Selling Formal Insurance to the InformallyInsured.” Tech. Rep. 1007, Economic Growth Center, Yale University.

———. 2014. “Risk, Insurance and Wages in General Equilibrium.” Working Paper 19811, NationalBureau of Economic Research.

Morduch, Jonathan. 1999. “The Microfinance Promise.” Journal of Economic Literature37 (4):1569–1614.

Navajas, Sergio, Mark Schreiner, Richard L. Meyer, Claudio Gonzalez-vega, and Jorge Rodriguez-meza. 2000. “Microcredit and the Poorest of the Poor: Theory and Evidence from Bolivia.”World Development 28 (2):333–346.

Quisumbing, Agnes R. 2007. “Poverty Transitions, Shocks, and Consumption in Rural Bangladesh:Preliminary Results from a Longitudinal Household Survey.” Tech. Rep. 105, Chronic PovertyResearch Centre Working Paper.

Rosenzweig, Mark R. and Hans P. Binswanger. 1993. “Wealth, Weather Risk and the Compositionand Profitability of Agricultural Investments.” The Economic Journal 103 (416):56–78.

Santos, Indhira, Iffath Sharif, Hossain Zillur Rahman, and Hassan Zaman. 2011. How Do the PoorCope with Shocks in Bangladesh? Evidence from Survey Data. Policy Research Working Papers.The World Bank.

Serfilippi, Elena, Michael Carter, and Catherine Guirkinger. 2018. “Insurance Contracts WhenIndividuals “Greatly Value” Certainty: Results from a Field Experiment in Burkina Faso.”Working Paper 25026, National Bureau of Economic Research.

Tedeschi, Gwendolyn Alexander. 2006. “Here Today, Gone Tomorrow: Can Dynamic IncentivesMake Microfinance More Flexible?” Journal of Development Economics 80 (1):84–105.

Udry, Christopher. 1994. “Risk and Insurance in a Rural Credit Market: An Empirical Investigationin Northern Nigeria.” The Review of Economic Studies 61 (3):495–526.

———. 1995. “Risk and Saving in Northern Nigeria.” The American Economic Review 85 (5):1287–1300.

UNDP. 2015. “MDG Bangladesh Progress Report.” Tech. rep., UNDP.

World Bank. 2016. “Bangladesh: Growing the Economy through Advances in Agri-culture.” http://www.worldbank.org/en/results/2016/10/07/bangladesh-growing-economy-through-advances-in-agriculture.

35


Figures

Figure 1: Loan Choices for Eligible Members

Eligible BRAC Member

Dabi Loan

Good LoanConditional on 100%

Repayment

Emergency LoanConditional on Flood

No Loan

Emergency LoanConditional on Flood

Notes: Figures shows a schematic representation of the loan choices facing a BRAC microfinance member. Thereare three types of loans: the normal Dabi loan, the Good Loan, and the Emergency Loan. The Good Loan is onlyavailable to borrowers who have taken a Dabi Loan and have made all on-time payments through the first sixmonths of the original loan. The offer of a Good Loan expires after two months. The Emergency Loan is onlyavailable after a flood has occurred, but it is offered whether or not the member currently has an active Dabi Loan.Members who take a Good Loan cannot also take an Emergency Loan and vice versa.

36


Figure 2: Map of Sample Branches

21

22

23

24

25

26

88 89 90 91 92

Longitude

Latit

ude

ControlTreated

Gauge

Notes: Map shows the locations of BRAC branches that participated in the experiment (triangles) as well as thewater level gauges used to monitor flood water levels (circles). Branches were selected based on their history offlooding and proximity to a water level gauge maintained by the Bangladeshi government.

37


Figure 3: Referral Slip

Referral Slip – Emergency Loan

Member Copy: Please keep Branch Name:…………………………………… Code: Branch contact #: Member Name:………………………………………………… Member No: VO Code: PO Name: Sign: Branch Manager Sign: If you have a completed form with a signature then you are guarenteed eligiblity for Emergency Loan Loan Conditions:

• River overflow and local area flooding confirmed by BRAC

Things to bring when getting Emergency Loan • Referral slip • Identification card

Loan Amount • Can take up to 50% of current or last

loan • Maximum of 50,000 taka

Ineligibility condition • If you take a Good Loan • Your branch area is not affected by

flooding - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Tear here - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

Referral Slip – Emergency Loan

Office Copy: Please keep

Branch Name:…………………………………… Code: Member contact #: Member Name:………………………………………………… Member No: VO Code: PO Sign: Branch Manager Sign: Accountant Sign:

†idv‡ij w¯øc - Bgv‡R©wÝ FY m`m¨ Kwct GwU msi¶b Kiyb †gqv`t 15/11/2016 ch©šÍ

eªv‡Âi bvgt…...…………………........…...…… †KvWt eªv‡Â †hvMv‡hv‡Mi bs

m`‡mï bvgt…...……………………………………………........ m`m¨ bst wfI †KvWt

wcIi bvgt ¯^v¶it eªvÂ g¨v‡bRv‡ii ¯v¶it

Avcbvi Kv‡Q hw` GB w¯¬cwU c~ibK…Z I g¨v‡bRv‡ii ¯^v¶imn _v‡K, Zvi gv‡b Avcwb Bgv‡R©wÝ F‡Yi Rb¨ wbe©vwPZ n‡q‡Qb|

FY cvIqvi kZ©vejxt x eªv‡Âi Kg© GjvKvq wbKU¯’ b`xi cvwb wec`mxgv AwZµg

K‡i eb¨vµvšÍ n‡q‡Q Zv wbwðZ n‡j|

F‡bi cwigvYt x PjwZ F‡Yi A_ev PjwZ FY bv _vK‡j me©‡kl F‡Yi

m‡e©v”P 50% ch©šÍ Bgv‡R©wÝ FY wb‡Z cvi‡eb x m‡e©v”P 50,000(cÂvk nvRvi UvKv) ch©šÍ Bgv‡R©wÝ FY

wb‡Z cvi‡eb

FY MÖn‡Yi mgq hv mv‡_ Avb‡Z n‡et x †idv‡ij w øc - m`m¨ Kwc x †fvUvi AvBwW KvW© / Rb¥wbeÜb KvW© x 1 Kwc cvm‡cvU© mvBR Qwe

FY cvIqvi †¶‡Î A‡hvM¨Zvt

x hw` ¸W FY Pjgvb _v‡K x Avcwb eb¨vq ¶wZMÖ¯’ n‡jI Avcbvi GjvKv eb¨vµvšÍ

bv n‡j

- - - - - - - - - - - - - - - - - - - - - - - - - -- - - - - - - - - -GLv‡b wQuo–b - - - - - - - - -- - - - - - - - - - - - - - - - - - - - - - - - - -

†idv‡ij w¯øc - Bgv‡R©wÝ FY Awdm Kwct GwU msi¶b Kiyb †gqv`t 15/11/2016 ch©šÍ

eªv‡Âi bvgt …...…………………........….. †KvWt m`‡mï †gvevBj bs


wcIi ¯^v¶it eªvÂ g¨v‡bRv‡ii ¯^v¶it eªvÂ A¨vKvD›U‡mi ¯^v¶it

†idv‡ij w¯øc - Bgv‡R©wÝ FY m`m¨ Kwct GwU msi¶b Kiyb †gqv`t 15/11/2016 ch©šÍ

eªv‡Âi bvgt…...…………………........…...…… †KvWt eªv‡Â †hvMv‡hv‡Mi bs


wcIi bvgt ¯^v¶it eªvÂ g¨v‡bRv‡ii ¯v¶it

Avcbvi Kv‡Q hw` GB w¯¬cwU c~ibK…Z I g¨v‡bRv‡ii ¯^v¶imn _v‡K, Zvi gv‡b Avcwb Bgv‡R©wÝ F‡Yi Rb¨ wbe©vwPZ n‡q‡Qb|

FY cvIqvi kZ©vejxt x eªv‡Âi Kg© GjvKvq wbKU¯’ b`xi cvwb wec`mxgv AwZµg

K‡i eb¨vµvšÍ n‡q‡Q Zv wbwðZ n‡j|

F‡bi cwigvYt x PjwZ F‡Yi A_ev PjwZ FY bv _vK‡j me©‡kl F‡Yi

m‡e©v”P 50% ch©šÍ Bgv‡R©wÝ FY wb‡Z cvi‡eb x m‡e©v”P 50,000(cÂvk nvRvi UvKv) ch©šÍ Bgv‡R©wÝ FY

wb‡Z cvi‡eb

FY MÖn‡Yi mgq hv mv‡_ Avb‡Z n‡et x †idv‡ij w øc - m`m¨ Kwc x †fvUvi AvBwW KvW© / Rb¥wbeÜb KvW© x 1 Kwc cvm‡cvU© mvBR Qwe

FY cvIqvi †¶‡Î A‡hvM¨Zvt

x hw` ¸W FY Pjgvb _v‡K x Avcwb eb¨vq ¶wZMÖ¯’ n‡jI Avcbvi GjvKv eb¨vµvšÍ

bv n‡j

- - - - - - - - - - - - - - - - - - - - - - - - - -- - - - - - - - - -GLv‡b wQuo–b - - - - - - - - -- - - - - - - - - - - - - - - - - - - - - - - - - -

†idv‡ij w¯øc - Bgv‡R©wÝ FY Awdm Kwct GwU msi¶b Kiyb †gqv`t 15/11/2016 ch©šÍ

eªv‡Âi bvgt …...…………………........….. †KvWt m`‡mï †gvevBj bs


wcIi ¯^v¶it eªvÂ g¨v‡bRv‡ii ¯^v¶it eªvÂ A¨vKvD›U‡mi ¯^v¶it

Notes: Shows the referral slip (translated from Bangla) given to BRAC microfinance members eligible for theEmergency Loan. The slip records clients name and BRAC identifiers, the maximum pre-approved loan size, as wellas a brief description of the loan product. The bottom of the slip also contained the borrower’s information and waskept by the branch manager to facilitate easy follow-up should a flood occur in the area.

38


Figure 4: BRAC Loans

Notes: Figure shows the uptake of the three different BRAC loan products examined in the experiment. The solidline shows Dabi loan uptake as a proportion of overall branch membership. The Short-dashed line shows GoodLoan uptake as a proportion of Good Loan eligible clients. The long-dashed line shows Emergency Loan uptake as aproportion of eligible clients offered the loan. The shaded regions show the Aman cropping season. The Good Loaneligibility data set is not usually recorded by BRAC, therefore there is a gap in this data between the 2016 and 2017Aman seasons when this data was not recorded because of uncertainty about the continuation of the experiment.

39


Figure 5: Yield Per Acre by Emergency Loan Uptake

0

.001

.002

.003

.004

Den

sity

0 2000 4000 6000 8000Yield per acre

EL Taker Non-Taker

Notes: Plots the yield per acre split by Emergency Loan takers and non-takers. Sample pools data from both 2016and 2017 and is limited to respondents who were Emergency Loan eligible and located in flooded branches.

40


Figure 6: Emergency Loan Uptake by Credit Score

●

●

●

●

●

●

●

●

●

●

0.00

0.02

0.04

0.06

78 80 82 84

Credit Score

Pro

babi

lity

Em

erge

ncy

Loan

Upt

ake

Notes: Plots the probability of Emergency Loan uptake by borrower credit score. Sample pools data from both2016 and 2017 and is limited to respondents who were Emergency Loan eligible and located in flooded branches.

41


0

2500

5000

7500

40 60 80 100

credit_score

coun

t

Number of Borrowers GL Eligible by EL Credit Score

Figure 7: Plots the distribution of Good Loan eligible borrowers across all treatment branches by their EmergencyLoan credit score. The eligibility threshold for the Emergency Loan (77) is shown in red.

42


Figure 8: Good Loan Uptake Heterogeneity

●

●

●

●

●

● ●

●

●

−0.04

−0.03

−0.02

−0.01

77 79 81 83

Credit Score

Trea

tmen

t Effe

ct o

n G

L U

ptak

e

Notes: Plots the treatment effect on the uptake of the Good Loan in treatment branches by decile of borrowercredit score. The regression run on each decile includes year and district fixed effects. Sample is comprised ofEmergency Loan eligible borrowers who were also eligible for a Good Loan in the pre-flood period. Standard errorsare clustered at the branch level. Table 18 tests whether the treatment effect heterogeneity is significant.

43


Figure 9: Dabi Loan Uptake Over Time

Notes: Plots the probability that an Emergency Loan eligible BRAC member takes a dabi loan in a given month.Probability of loan uptake is calculated using the complete number of BRAC members in each branch, regardless ofwhether or not they have a current dabi loan. This is to ensure that the denominator does not endogenously changebased on pervious loan uptake decisions. The shaded regions are the “pre-period” before the beginning of the floodseason in 2016 and 2017. This graph corresponds to regression table 16.

44


Figure 10: Dabi Loan Uptake Heterogeneity

●

●

●

●

●

●●

●

●

0.000

0.005

0.010

0.015

78 80 82 84 86

Credit Score

Trea

tmen

t Effe

ct o

n D

abi U

ptak

e

Notes: Plots the treatment effect on the uptake of the Dabi Loan by decile of borrower credit score. The regressionrun on each decile includes year, month, and district fixed effects. The sample includes only Emergency Loaneligible borrowers. Standard errors are clustered at the branch level. Table 18 tests whether the treatment effectheterogeneity is significant.

45


Figure 11: Missed Payment Treatment Effect Heterogeneity

●

●

●

●

●

●

●

●

●

−0.05

−0.04

−0.03

−0.02

−0.01

77.5 80.0 82.5 85.0

Credit Score

Trea

tmen

t Effe

ct o

n P

roba

bilit

y of

Mis

sed

Pay

men

t

Notes: Plots the treatment effect on the probability of a missed payment by decile of borrower credit score. Theestimated treatment effect is the average change in repayment rate across both flooded and non-flooded branches.The regression run on each decile includes year, month, and district fixed effects. The sample includes onlyEmergency Loan eligible borrowers. Standard errors are clustered at the branch level. Table 18 tests whether thetreatment effect heterogeneity is significant.

Figure 12: Missed Payment Heterogeneity

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●●

●

●

0.0

0.1

0.2

0.3

78 80 82 84

Credit Score

Pro

babi

lity

Mis

s P

aym

ent

Treatment●

●

0

1

Notes: Plots the probability of a missed payment by decile of borrower credit score separately for treatment andcontrol branches. The sample is comprised of only Emergency Loan eligible borrowers.

46


Figure 13: Per-Person Profits Heterogeneity

●●

●●

●●

●●

●●

●●

●●

●●

●●

−500

−250

0

250

77 78 79 80 81 82 83

Credit Score

Trea

tmen

t Effe

ct o

n P

er P

erso

n P

rofit

s

Notes: Plots the treatment effect on per-person profits by decile of borrower credit score. Profits are measured inBangladeshi taka ($1 = 84tk). The sample includes only Emergency Loan eligible borrowers. Standard errors areclustered at the branch level. Table 18 tests whether the treatment effect heterogeneity is significant.

47


Table 1: Eligible Compared to Ineligible

(1) (2) (3)Ineligible Eligible p-value of equality

Household Size 4.788 4.893 0.010(0.030) (0.027)

Age Head of Household 39.831 40.763 0.004(0.246) (0.208)

Educ. Head of Household 2.772 2.497 0.001(0.069) (0.053)

Acres of Land Owned 0.461 0.454 0.868(0.021) (0.032)

Household Income 1627.133 1560.817 0.042(26.429) (20.100)

Weekly Expenditure 22.256 22.330 0.873(0.344) (0.305)

Flooded in Past 0.537 0.543 0.598(0.009) (0.007)

Electricity Access 0.706 0.717 0.265(0.008) (0.007)

Asset Count 1.659 1.678 0.418(0.018) (0.015)

Cows Owned 0.741 0.916 0.000(0.023) (0.021)

Risk Aversion 0.499 0.513 0.147(0.007) (0.006)

Notes: Table compares households that were eligible for the Emergency Loan to those who were ineligible in bothtreatment and control branches at baseline conducted in April 2016. Asset count is the number of items ahousehold reported owning of a gas or electric stove, radio, television, refrigerator, bicycle, and motorcycle. Riskaversion was measured by asking households to choose between a certain payoff and a lottery with increasing odds.The measure ranges from zero to one, where 0=most risk loving and 1=most risk averse.

48


Table 2: Research Timeline

Oct 2015 - Jan 2016 · · ·• Development of product.

Feb 2016 · · ·• 200 experimental branches selected.

Apr 2016 · · ·•Baseline survey of 4,000 households; Year onecredit scores created; Clients informed abouteligibility.

Jun - Oct 2016 · · ·• Flood monitoring and Emergency Loans madeavailable as necessary.

Dec 2016 · · ·• Follow-up survey of 4,000 households.

Apr 2017 · · ·• Year two credit scores created; Clientsinformed about eligibility.

Jun - Oct 2017 · · ·• Flood monitoring and Emergency Loans madeavailable as necessary.

Dec 2017 · · ·• Endline survey of 4,000 households.

49


Table 3: Balance Table

(1) (2) (3)Control Treatment p-value of equality

test

Household Size 4.867 4.874 0.910(0.047) (0.046)

Age Head of Household 40.883 40.374 0.339(0.371) (0.381)

Educ. Head of Household 2.542 2.464 0.564(0.095) (0.095)

Acres of Land Owned 0.394 0.436 0.202(0.021) (0.025)

Household Income 1594.585 1537.005 0.244(34.486) (35.453)

Weekly Expenditure 21.989 22.191 0.779(0.485) (0.531)

Flooded in Past 0.527 0.548 0.250(0.013) (0.013)

Electricity Access 0.707 0.724 0.326(0.012) (0.012)

Asset Count 1.724 1.658 0.076(0.026) (0.027)

Cows Owned 0.887 0.922 0.497(0.035) (0.039)

Risk Aversion 0.509 0.511 0.905(0.010) (0.010)

Notes: Table compares households in treatment and control branches at baseline conducted in April 2016 beforetreatment status was revealed. Asset count is the number of items a household reported owning of a gas or electricstove, radio, television, refrigerator, bicycle, and motorcycle. Risk aversion was measured by asking households tochoose between a certain payoff and a lottery with increasing odds. The measure ranges from zero to one, where0=most risk loving and 1=most risk averse.

50


Table 4: Flood Summary

2016Treatment Flooded Branches

No No 60No Yes 40Yes No 49Yes Yes 51

2017Treatment Flooded Branches

No No 27No Yes 73Yes No 37Yes Yes 63

51


Table 5: Emergency Loan Uptake

(1) (2)Took Emergency Loan Took Emergency Loan

Baseline HH Income -0.005(0.003)

Risk Aversion 0.007(0.013)

Baseline Time -0.003Preference (0.002)

Number of Past -0.008Floods (0.005)

Ex-post Investment 0.021Opportunity (0.016)

Preparation for -0.026∗

flood (1=low, 5=high) (0.014)

Distress from flood 0.054∗∗∗

(1=low, 5=high) (0.014)

Controls Yes Yes

District FE Yes Yes

Mean Dep. Var 0.03 0.05Observations 1193 525

Notes: Sample includes only treatment BRAC members who were eligible to take an Emergency Loan in anactivated branch. The outcome variable is an indicator for the borrower taking the offered Emergency Loan.Standard errors clustered at branch level. Column 1 shows results predicting Emergency Loan take-up using datacollected at baseline. Yearly household income is measured in thousands of dollars. Risk aversion ranges 0 to 1,where 0=most risk loving and 1=most risk averse. Time preference ranges from 1 to 9, where 1 = most impatientand 9 = most patient. Number of past floods is the number of flood shocks experienced by the household over theprevious five years (2011-2016). Column 2 predicts Emergency Loan take-up using data gathered at endline andonly has observations from 2017. Flood preparation was measured at baseline. Ex-post investment opportunity isan indicator for whether the household reported having a good investment opportunity after the flood. Preparationfor flood and distress from flood were self-reported by households.

52


Table 6: Uptake of Good Loan by Emergency Loan Availability

Took Good Loan

Treatment −0.020∗∗ −0.022∗∗ −0.020∗∗

(0.008) (0.009) (0.008)

Farming x Treatment 0.006(0.016)

Farming Main Activity −0.007(0.010)

Flood Risk x Treatment −0.015∗∗∗

(0.006)

Flood Risk 0.011∗∗∗

(0.004)

Year F.E. Yes Yes YesDistrict F.E. Yes Yes YesMean of Dependent Var 0.130 0.130 0.129Unique Borrowers 66,232 66,232 63,744Observations 75,818 75,818 73,282

Notes: Sample is comprised of Good Loan eligible clients who were offered a Good Loan in the pre-flood period.Observations at the month-person level. Data is pooled from both 2016 and 2017. Standard errors clustered atbranch level. The outcome variable is an indicator for whether or not the borrower took the offered Good Loan.Farming is a branch level indicator for farming being the major source of income for BRAC members in thatbranch. Flood risk is measured at the branch level on 1-5 scale where 1 = least risk and 5 = high risk.

53


Table 7: Land Farmed

(1) (2) (3) (4) (5)Own land Rented land Sharecrop land Total land Any Cult.

Treatment 0.000 0.063∗∗∗ -0.004 0.058∗∗ 0.044∗

(0.013) (0.016) (0.004) (0.026) (0.024)

Controls Yes Yes Yes Yes Yes

District FE Yes Yes Yes Yes Yes

Mean Dep. Var 0.13 0.20 0.02 0.35 0.46Observations 4744 4740 4743 4739 4745

Notes: Sample includes only eligible BRAC members from both treatment and control groups. Data is pooledfrom both the 2016 and 2017 Aman season. Controls are included for precision, and are comprised of baselinemeasures of total land owned, household size, and the age and education of the head of household. Standard errorsclustered at the branch level. Land measured in acres. Total land is the sum of own land, rented land, andsharecropped land. Any Cult. is an indicator for whether or not a household planted any crops during the season.

Table 8: Ex-Ante Investments

(1) (2) (3) (4) (5)Fert. Applied Pest. Applied Cost Seeds per acre Input Cost per Acre Non-Ag Invest

Treatment 6.51 0.26 0.32 2.06 12.13∗

(5.30) (0.17) (0.76) (2.17) (6.64)




Notes: Sample includes only eligible BRAC members from both treatment and control groups. Data is pooledfrom both the 2016 and 2017 Aman season. Controls are included for precision, and are comprised of baselinemeasures of total land owned, household size, and the age and education of the head of household. Standard errorsclustered at the branch level. Fertilizer and pesticide measured in kg/L per acre. Input cost per acre is the sum ofthe cost of fertilizer, pesticide, and seeds. Cost and investment measured in dollars.

54


Tab

le9:

Ex-A

nte

Lan

dby

Ris

kA

vers

ion

(1)

(2)

(3)

(4)

(5)

Ow

nla

nd

Ren

ted

land

Share

crop

land

Tota

lla

nd

Any

Cult

.

Tre

atm

ent

-0.0

14

0.0

35

-0.0

07

0.0

07

0.0

37

(0.0

21)

(0.0

25)

(0.0

06)

(0.0

36)

(0.0

31)

Ris

kA

ver

sion

X0.0

20

0.0

61∗

0.0

06

0.0

97∗∗

0.0

13

Tre

atm

ent

(0.0

31)

(0.0

36)

(0.0

09)

(0.0

49)

(0.0

41)

Ris

kA

ver

sion

0.1

82∗∗

-0.0

03

-0.0

08

0.1

63∗

0.0

75

(0.0

71)

(0.0

53)

(0.0

11)

(0.0

89)

(0.0

78)

Contr

ols

Yes

Yes

Yes

Yes

Yes

Contr

ols

XR

isk

Aver

sion

Yes

Yes

Yes

Yes

Yes

Dis

tric

tF

EY

esY

esY

esY

esY

es

Mea

nD

ep.

Var

0.1

30.2

00.0

20.3

60.4

7O

bse

rvati

ons

4479

4475

4478

4474

4480

p-v

alu

eT

reat

+R

isk

XT

reat

0.7

56

0.0

00

0.8

30

0.0

04

0.1

31

Note

s:Sam

ple

incl

udes

only

elig

ible

BR

AC

mem

ber

sfr

om

both

trea

tmen

tand

contr

ol

gro

ups.

Data

isp

oole

dfr

om

both

the

2016

and

2017

Am

an

seaso

n.

Contr

ols

are

incl

uded

for

pre

cisi

on,

and

are

com

pri

sed

of

base

line

mea

sure

sof

tota

lla

nd

owned

,house

hold

size

,and

the

age

and

educa

tion

of

the

hea

dof

house

hold

.Sta

ndard

erro

rscl

ust

ered

at

bra

nch

level

.L

and

ism

easu

red

inacr

es.

Tota

lla

nd

isth

esu

mof

own

land,

rente

dla

nd,

and

share

cropp

edla

nd.

Any

Cult

.is

an

indic

ato

rfo

rw

het

her

or

not

ahouse

hold

pla

nte

dany

crops

duri

ng

the

seaso

n.

Ris

kav

ersi

on

was

mea

sure

dat

base

line

and

ranges

0to

1,

wher

e0=

most

risk

lovin

gand

1=

most

risk

aver

se.

55


Tab

le10

:E

x-A

nte

Inp

uts

by

Ris

kA

vers

ion

(1)

(2)

(3)

(4)

(5)

Fer

t.A

pplied

Pes

t.A

pplied

Cost

See

ds

per

acr

eIn

put

Cost

per

Acr

eN

on-A

gIn

ves

t

Tre

atm

ent

6.4

40.0

51.1

21.6

83.4

4(7

.80)

(0.3

0)

(1.2

4)

(3.7

7)

(11.7

7)

Ris

kA

ver

sion

X1.6

40.4

1-1

.34

0.6

516.0

6T

reatm

ent

(13.1

8)

(0.4

3)

(1.7

8)

(5.4

1)

(16.6

2)

Ris

kA

ver

sion

2.3

1-0

.96

-4.9

5-1

7.6

1∗

17.3

1(2

3.9

3)

(0.7

9)

(3.6

5)

(10.1

8)

(32.2

5)

Contr

ols

Yes

Yes

Yes

Yes

Yes

Contr

ols

XR

isk

Aver

sion

Yes

Yes

Yes

Yes

Yes

Dis

tric

tF

EY

esY

esY

esY

esY

es

Mea

nD

ep.

Var

138.7

11.5

316.0

865.5

033.0

8O

bse

rvati

ons

2089

2048

1971

1932

4480

p-v

alu

eT

reat

+R

isk

XT

reat

0.3

58

0.0

60

0.8

33

0.4

63

0.0

28

Note

s:Sam

ple

incl

udes

only

elig

ible

BR

AC

mem

ber

sfr

om

both

trea

tmen

tand

contr

ol

gro

ups.

Data

isp

oole

dfr

om

both

the

2016

and

2017

Am

an

seaso

n.

Contr

ols

are

incl

uded

for

pre

cisi

on,

and

are

com

pri

sed

of

base

line

mea

sure

sof

tota

lla

nd

owned

,house

hold

size

,and

the

age

and

educa

tion

of

the

hea

dof

house

hold

.Sta

ndard

erro

rscl

ust

ered

at

bra

nch

level

.F

erti

lize

rand

pes

tici

de

are

mea

sure

din

kg

/L

per

acr

e.In

put

cost

per

acr

eis

the

sum

of

the

cost

of

fert

iliz

er,

pes

tici

de,

and

seed

s.C

ost

and

inves

tmen

tm

easu

red

indollars

.R

isk

aver

sion

ranges

0to

1,

wher

e0=

most

risk

lovin

gand

1=

most

risk

aver

se.

56


Table 11: Investment After Shock

(1) (2) (3) (4) (5)Fert. Applied Pest. Applied Total land Any Cult. Non-Ag Invest

Treatment 6.689 0.323∗ 0.055∗∗ 0.035 12.559∗

(5.795) (0.192) (0.028) (0.025) (6.397)

Flood Last Year X 0.053 -0.339 0.021 0.063 0.358Treat (23.333) (0.556) (0.044) (0.046) (24.457)

Flood Last Year -4.615 -0.383 -0.033 -0.099∗∗ -21.348(20.213) (0.488) (0.042) (0.045) (23.778)



Mean Dep. Var 140.47 1.58 0.35 0.46 38.69Observations 2183 2140 4739 4745 4745p-value Treat + Interaction 0.757 0.974 0.069 0.029 0.591

Notes: Sample includes only eligible BRAC members from both treatment and control groups. Data is pooledfrom both the 2016 and 2017 Aman season. Controls are included for precision, and are comprised of baselinemeasures of total land owned, household size, and the age and education of the head of household. Standard errorsclustered at the branch level. Fertilizer and pesticide measured in kg/L per acre. Total land is the sum of own land,rented land, and sharecropped land. Any Cult. is an indicator for whether or not a household planted any cropsduring the season. Investment is measured in dollars.

57


Table 12: Ex-Post Outcomes

(1) (2) (3) (4)Log Cons PerCap Log Income Crop Prod. (Kg) Livestock

Treatment 0.050 -0.024 92.104∗∗ -0.075(0.046) (0.044) (41.259) (0.106)

Flood X Treatment 0.058 0.002 -83.157 0.353∗∗

(0.062) (0.063) (51.968) (0.144)

Flood -0.046 0.030 -0.831 0.058(0.059) (0.057) (38.074) (0.109)

Controls Yes Yes Yes Yes

District FE Yes Yes Yes Yes

Week Interviewed FE Yes No No No

Mean Dep. Var 5.92 10.77 277.07 1.51Observations 4699 4489 4701 4701p-value Treat + Flood X Treat 0.011 0.609 0.800 0.007

Notes: Sample includes only eligible BRAC members from both treatment and control groups. Data is pooled fromboth the 2016 and 2017 Aman season. Controls are included for precision, and are comprised of baseline measuresof total land owned, household size, and the age and education of the head of household. Week interviewed fixedeffects are included for the log consumption regression due to the presence of holidays over the course of the surveyperiod that changed standard consumption patterns. Standard errors clustered at branch level. Income is measuredin dollars. Flood is an indicator that equals one if flooding occurred and the Emergency Loan was activated.

58


Table 13: Ex-post After Successive Shocks


Treatment 0.036 -0.023 93.639∗∗ -0.083(0.046) (0.044) (41.287) (0.107)

Flood X Treatment 0.107 -0.003 -99.495∗ 0.379∗∗

(0.067) (0.072) (54.868) (0.146)

Flood Current Year -0.051 0.032 5.382 0.056(0.059) (0.060) (38.331) (0.108)

Flood Both X Treat -0.100 0.017 54.321 -0.055(0.095) (0.096) (44.995) (0.171)

Flood Both Years -0.199∗∗∗ -0.000 -0.260 -0.100(0.069) (0.072) (41.944) (0.131)




Mean Dep. Var 5.92 10.77 277.07 1.51Observations 4699 4489 4701 4701p-value Sum Treatment Coef. 0.004 0.904 0.229 0.161

Notes: Sample includes only eligible BRAC members from both treatment and control groups. Data is pooledfrom both the 2016 and 2017 Aman season. Controls are included for precision, and are comprised of baselinemeasures of total land owned, household size, and the age and education of the head of household. Weekinterviewed fixed effects are included for the log consumption regression due to the presence of holidays over thecourse of the survey period that changed standard consumption patterns. Standard errors clustered at branch level.Income is measured in dollars. Flood Current Year is an indicator that equals one if flooding occurred in thecurrent year. Flood both years is an indicator that captures the additional effect of successive shocks for branchesthat experienced flooding in 2017 that also experienced flooding in 2016.

59


Table 14: Day Labor

(1) (2)Daily Wage Days Worked

Treatment 0.45 -1.57∗∗

(9.77) (0.78)

Flood X Treatment -0.90 2.02∗

(11.76) (1.17)

Flood 0.91 -4.50∗∗∗

(8.40) (0.94)

Controls Yes Yes

District FE Yes Yes

Mean Dep. Var 306.75 14.79Observations 928 2776p-value Treat + Flood X Treat 0.941 0.573

Notes: Sample includes only eligible BRAC members from both treatment and control groups. Data is pooledfrom both the 2016 and 2017 Aman season. Controls are included for precision, and are comprised of baselinemeasures of total land owned, household size, and the age and education of the head of household. Standard errorsclustered at branch level. Daily wage is in taka per day. Days worked is the number of days working for othersproviding farm labor or construction (construction was often for agriculture related infrastructure such as repairingirrigation channels).

60


Table 15: Dabi Loan Uptake by Emergency Loan Availability

Loan Uptake

Treatment 0.007∗∗∗

(0.002)

Year & Month F.E. YesDistrict F.E. YesMean of Dep. Var. 0.062Unique Borrowers 108,446Observations 462,172

Notes: Sample is comprised of all Emergency Loan eligible clients in the pre-flood period. Observations at themonth-person level. Data is pooled from both the 2016 and 2017. Standard errors clustered at branch level. Theoutcome variable is an indicator for whether or not the client took a new dabi loan in the period before the floodseason.

61


Table 16: Repayment by Emergency Loan Availability

Missed Payment

Treatment 0.011(0.024)

Treat x Flood −0.040∗

(0.020)

Flood 0.039∗

(0.023)

Year & Month F.E. YesDistrict F.E. YesMean of Dep. Var. 0.096Unique Borrowers 109,647Observations 378,216

Notes: Sample includes only Emergency Loan eligible clients. Standard errors clustered at branch level.Observations at the loan-month level. The outcome variable is an indicator for whether or not the client missed aloan payment in a given month. The variable flood is an indicator for anytime after a flood until the followingMarch.

62


Table 17: Branch Profit by Emergency Loan Availability

Profit (Taka)Per Loan Monthly Branch Monthly Per Person

(1) (2) (3)

Treatment 161 76,312 96∗∗

(233) (95,405) (46)

District F.E. Yes Yes YesMonth F.E. No Yes YesMean of Dep. Var. 2,823 1,745,794 2202Observations 106,695 3,706 3,706

Notes: Sample for column 1 includes loans made only to Emergency Loan eligible clients. Standard errorsclustered at branch level. The outcome for column 1 is the measured profit in Bangladeshi taka ($1 = 84 taka) for agiven loan assuming an annual cost of capital of 6% for the MFI. The outcome for column 2 is overall branchprofitability. The outcome in column 3 is overall branch profitability divided by the number of branch members.Observations in column 1 are at the loan level and for column 2 and 3 are at the branch-month level.

63


Tab

le18

:M

FI

Ou

tcom

esby

Cre

dit

Sco

re

Good

Loa

nU

pta

keD

abi

Up

take

Mis

sed

Pay

men

tP

erP

erso

nP

rofi

t

(1)

(2)

(3)

(4)

Tre

atm

ent

−0.

020∗

0.00

8∗∗∗

−0.

027∗∗

169∗∗

(0.0

11)

(0.0

02)

(0.0

13)

(78.

034)

Cre

dit

Sco

rex

Tre

atm

ent

−0.

003∗

0.00

00.

004∗

−25∗

(0.0

02)

(0.0

002)

(0.0

02)

(14.

7)

Cre

dit

Sco

re0.

004∗∗∗

−0.

0001

−0.

010∗∗∗

13∗∗

(0.0

01)

(0.0

002)

(0.0

02)

(5.7

40)

Dis

tric

tF

.E.

Yes

Yes

Yes

Yes

Mon

thF

.E.

No

Yes

Yes

No

Yea

rF

.E.

Yes

Yes

Yes

No

Mea

nof

Dep

.V

ar.

0.13

0.06

20.

096

2202

Ob

serv

ati

on

s37

,392

396,

228

910,

862

40,5

14

Note

s:Sam

ple

incl

udes

only

Em

ergen

cyL

oan

elig

ible

clie

nts

.Sta

ndard

erro

rscl

ust

ered

at

bra

nch

level

.T

he

outc

om

ein

colu

mn

1is

the

pro

babilit

yof

takin

gan

off

ered

Good

Loan

am

ong

Good

Loan

elig

ilb

ecl

ients

inth

epre

-flood

per

iod.

The

outc

om

ein

colu

mn

2is

the

pro

babilit

yof

takin

ga

Dabi

Loan

inth

epre

-flood

per

iod.

The

outc

om

ein

colu

mn

3is

the

pro

babilit

yof

mis

sing

alo

an

pay

men

tin

agiv

enm

onth

.T

he

outc

om

ein

colu

mn

4is

the

mea

sure

dpro

fit

inB

angla

des

hi

taka

per

bra

nch

mem

ber

ass

um

ing

an

annual

cost

of

capit

al

of

6%

for

the

MF

I.

64


Table 19: Savings Transactions by Emergency Loan Availability

Savings TransctionsPre-Period All All (2017)

(1) (2) (3)

Treatment 8.85 −14.58 −55.73(9.34) (18.57) (43.11)

Treat x Flood 45.37∗∗ 34.75∗

(20.67) (20.75)

Flood −53.75∗∗ −50.19∗∗

(24.60) (22.19)

Flood Damage x Treatment 11.58(10.05)

Flood Damage −17.15∗∗∗

(6.42)

Year & Month F.E. Yes Yes YesDistrict F.E. Yes Yes YesMean of Dep. Var. 82.6 71.8 64.5Unique Accounts 108,446 109,647 75,477Observations 622,551 1,150,895 711,184

Notes: Sample includes only Emergency Loan eligible clients. Standard errors clustered at branch level.Observations at the person-month level. The variable flood is an indicator for anytime after a flood until thefollowing March. Column 1 uses observations only from the pre-flood period in both 2016 and 2017. Column 2 usesall observations. Flood damage data at the branch level is only available for 2017, therefore column 3 shows resultsonly for this year. Flood damage is measured at the branch level and ranges from [1-5] with 1=least damage and5=most damage.

65


Appendix A: Comparative Statics

In this section we will more formally derive the comparative statics for input choice x and firstperiod borrowing b1 with respect to the increase in second period borrowing b2B. Starting with themaximization problem defined in equation 9:

maxx,b1,b2B

L = u(Y − x+ b1) + qβu(−Rb1 + b2B

)+ (1− q)βu

(mGf(x)−Rb1

)+

qβ2u(I −Rb2B) + (1− q)β2u(I) + λ1[B1 − b1] + λ2[B2 − b2B]

Where the FOCs are given by:

∂L∂x

=− u′(c1) + (1− q)βu′(c2G)mGf′

∂L∂b1

=u′(c1)− qβRu′(c2B)− (1− q)βRu′(c2g)− λ1∂L∂b2B

=qβu′(c2B)− qRβ2u′(c3B)− λ2

Note, we assume the constraints do not bind (λt = 0) so that the choice of x and b1 can adjust.

We also know from the implicit function theory that we can calculate ∂x∂b2B

and ∂b1

∂b2Bby:[ ∂x

∂b2B∂b1

∂b2B

]= −

[∂L

∂x∂x∂L

∂x∂b1∂L

∂b1∂x∂L

∂b1∂b1

]−1 [ ∂L∂x∂b2B∂L

∂b1∂b2B

]Calculating each term separately:

∂L∂x∂x

= u′′(c1) + (1− q)βmG

[(f ′)2u′′(c2G) + f ′′u′(c2G)

]< 0

∂L∂x∂b1

= −u′′(c1)− qβRmGf′u′′(c2G) > 0

∂L∂b1∂x

= −u′′(c1)− qβRmGf′u′′(c2G) > 0

∂L∂b1∂b1

= u′′(c1) + βR2[qu′′(c2B) + (1− q)u′′(c2G)

]< 0

∂L∂x∂b2B

= 0

∂L∂b1∂b2B

= −qβRu′′(c2B) > 0

Inverting the matrix[ ∂x∂b2B∂b1

∂b2B

]= − 1

∂L∂x∂x

∂L∂b1∂b1

− ∂L∂x∂b1

∂L∂b1∂x

[∂L

∂b1∂b1− ∂L

∂x∂b1

− ∂L∂b1∂x

∂L∂x∂x

][ ∂L∂x∂b2B∂L

∂b1∂b2B

]The denominator of the fraction is the determinate of a 2x2 hessian from a maximization problem,

66


and is therefore positive. Then, the matrices are pre-multiplied by a negative value, which we willreplace with − 1

Det . Multiplying out the matrices we find

∂x

∂b2B= − 1

Det︸︷︷︸−

[∂L

∂b1∂b1· 0− ∂L

∂x∂b1∂L

∂b1∂b2B

]︸︷︷︸

−

> 0

∂b1

∂b2B= − 1

Det︸︷︷︸−

[− ∂L∂b1∂x

· 0 +∂L∂x∂x

∂L∂b1∂b2B

]︸︷︷︸

−

> 0

Therefore, we conclude that the choice of inputs x and first period borrowing b1 will both increasewith the offer of the Emergency Loan.

67


Appendix B

In this appendix we examine whether selection into eligibility in 2017 matters for the results.First, we simply examine whether there was differential Emergency Loan eligibility in 2017 acrosstreatment and control branches. We see in Table 20 shows that there is no statistically signifi-cant difference in the probability that households are Emergency Loan eligible between treatmentand control branches. Ignoring statistical significance, the point estimate suggests that treatmentbranches were three percentage points less likely to be Emergency Loan eligible in 2017. This is theopposite effect as what might be expected ex-ante, that households in treatment branches improverepayment rates and are therefore more likely to become eligible. Finally, I also report ex-postoutcomes without controlling for flooding.

Table 20: 2017 Eligiblity

(1)EL Eligible

Treatment Branch -0.030(0.029)

Flood Last Year Yes

District FE Yes

Observations 3939

Notes: Sample includes all surveyed households in 2017. The outcome variable is a binary indicator for thehousehold being Emergency Loan eligible in 2017. Flood last year is an indicator for being flooded in 2016.

As a robustness check, I reproduce the results on household investment and ex-post outcomes withtwo different specifications. First, I limit the analysis to only 2016 when there are no selectionconcerns. Second, I instrument for eligibility using branch treatment status. With the exception ofnon-agriculture investment, the results are consistent with those found with the other specifications.

68


Table 21: Land Farmed 2016


Treatment 0.001 0.067∗∗∗ -0.006 0.059∗ 0.034(0.014) (0.020) (0.004) (0.030) (0.027)




Notes: Sample includes only eligible BRAC members from both treatment and control groups. Data is from onlythe 2016 Aman season. Controls are included for precision, and are comprised of baseline measures of total landowned, household size, and the age and education of the head of household. Standard errors clustered at the branchlevel. Land measured in acres. Total land is the sum of own land, rented land, and sharecropped land. Any Cult. isan indicator for whether or not a household planted any crops during the season.

Table 22: Ex-Ante Investments 2016


Treatment 6.15 0.36∗ 1.05 1.20 1.09(5.62) (0.18) (0.89) (2.49) (3.35)




Notes: Sample includes only eligible BRAC members from both treatment and control groups. Data is only fromthe 2016 Aman season. Controls are included for precision, and are comprised of baseline measures of total landowned, household size, and the age and education of the head of household. Standard errors clustered at the branchlevel. Fertilizer and pesticide measured in kg/L per acre. Input cost per acre is the sum of the cost of fertilizer,pesticide, and seeds. Cost and investment measured in dollars.

69


Table 23: IV Land Farmed


Treatment -0.004 0.071∗∗∗ -0.007∗ 0.057∗ 0.034(0.015) (0.019) (0.004) (0.029) (0.028)




Notes: Sample includes all observations from both treatment and control groups. Treatment is instrumented usingfirst year eligibility interacted by year. Data is pooled from both the 2016 and 2017 Aman season. Controls areincluded for precision, and are comprised of baseline measures of total land owned, household size, and the age andeducation of the head of household. Standard errors clustered at the branch level. Land measured in acres. Totalland is the sum of own land, rented land, and sharecropped land. Any Cult. is an indicator for whether or not ahousehold planted any crops during the season.

Table 24: IV Inputs


Treatment 5.71 0.28 0.39 1.79 1.15(5.41) (0.18) (0.83) (2.38) (7.51)




Notes: Sample includes all observations from both treatment and control groups. Treatment is instrumented usingfirst year eligibility interacted by year. Data is pooled from both the 2016 and 2017 Aman season. Controls areincluded for precision, and are comprised of baseline measures of total land owned, household size, and the age andeducation of the head of household. Standard errors clustered at the branch level. Land measured in acres. Totalland is the sum of own land, rented land, and sharecropped land. Any Cult. is an indicator for whether or not ahousehold planted any crops during the season.

70


Tab

le25

:E

x-A

nte

Lan

dby

Ris

kA

vers

ion

:20

16

(1)

(2)

(3)

(4)

(5)

Ow

nla

nd

Ren

ted

land

Share

crop

land

Tota

lla

nd

Any

Cult

.

Tre

atm

ent

-0.0

20.0

4-0

.01∗

-0.0

00.0

3(0

.02)

(0.0

3)

(0.0

1)

(0.0

4)

(0.0

3)

Ris

kA

ver

sion

X0.0

30.0

50.0

10.1

1∗

0.0

0T

reatm

ent

(0.0

3)

(0.0

5)

(0.0

1)

(0.0

6)

(0.0

5)

Ris

kA

ver

sion

0.1

4∗

0.0

9-0

.02

0.1

9∗

0.1

2(0

.08)

(0.0

6)

(0.0

1)

(0.1

0)

(0.0

9)

Contr

ols

Yes

Yes

Yes

Yes

Yes

Contr

ols

XR

isk

Aver

sion

Yes

Yes

Yes

Yes

Yes

Dis

tric

tF

EY

esY

esY

esY

esY

es

Mea

nD

ep.

Var

0.1

50.2

20.0

20.3

90.5

0O

bse

rvati

ons

2986

2986

2986

2986

2986

p-v

alu

eT

reat

+R

isk

XT

reat

0.6

54

0.0

01

0.9

00

0.0

08

0.3

52

Note

s:Sam

ple

incl

udes

only

elig

ible

BR

AC

mem

ber

sfr

om

both

trea

tmen

tand

contr

ol

gro

ups.

Data

isonly

from

the

2016

Am

an

seaso

n.

Contr

ols

are

incl

uded

for

pre

cisi

on,

and

are

com

pri

sed

of

base

line

mea

sure

sof

tota

lla

nd

owned

,house

hold

size

,and

the

age

and

educa

tion

of

the

hea

dof

house

hold

.Sta

ndard

erro

rscl

ust

ered

at

bra

nch

level

.L

and

ism

easu

red

inacr

es.

Tota

lla

nd

isth

esu

mof

own

land,

rente

dla

nd,

and

share

cropp

edla

nd.

Any

Cult

.is

an

indic

ato

rfo

rw

het

her

or

not

ahouse

hold

pla

nte

dany

crops

duri

ng

the

seaso

n.

Ris

kav

ersi

on

was

mea

sure

dat

base

line

and

ranges

0to

1,

wher

e0=

most

risk

lovin

gand

1=

most

risk

aver

se.

71


Tab

le26

:E

x-A

nte

Inp

uts

by

Ris

kA

vers

ion

:20

16

(1)

(2)

(3)

(4)

(5)

Fer

t.A

pplied

Pes

t.A

pplied

Cost

See

ds

per

acr

eIn

put

Cost

per

Acr

eN

on-A

gIn

ves

t

Tre

atm

ent

7.4

00.2

51.8

52.0

2-0

.77

(9.9

8)

(0.3

0)

(1.4

4)

(4.4

3)

(5.3

3)

Ris

kA

ver

sion

X-3

.29

0.2

0-1

.57

-1.8

93.5

8T

reatm

ent

(16.8

7)

(0.4

4)

(1.8

8)

(6.5

2)

(8.2

9)

Ris

kA

ver

sion

9.3

30.3

1-4

.80

-5.6

5-8

.93

(28.4

1)

(0.8

1)

(3.5

6)

(12.5

6)

(13.7

7)

Contr

ols

Yes

Yes

Yes

Yes

Yes

Contr

ols

XR

isk

Aver

sion

Yes

Yes

Yes

Yes

Yes

Dis

tric

tF

EY

esY

esY

esY

esY

es

Mea

nD

ep.

Var

129.9

31.3

414.4

560.5

37.8

4O

bse

rvati

ons

1479

1479

1375

1375

2986

p-v

alu

eT

reat

+R

isk

XT

reat

0.6

89

0.1

01

0.8

08

0.9

71

0.6

00

Note

s:Sam

ple

incl

udes

only

elig

ible

BR

AC

mem

ber

sfr

om

both

trea

tmen

tand

contr

ol

gro

ups.

Data

isonly

from

the

2016

Am

an

seaso

n.

Contr

ols

are

incl

uded

for

pre

cisi

on,

and

are

com

pri

sed

of

base

line

mea

sure

sof

tota

lla

nd

owned

,house

hold

size

,and

the

age

and

educa

tion

of

the

hea

dof

house

hold

.Sta

ndard

erro

rscl

ust

ered

at

bra

nch

level

.F

erti

lize

rand

pes

tici

de

are

mea

sure

din

kg

/L

per

acr

e.In

put

cost

per

acr

eis

the

sum

of

the

cost

of

fert

iliz

er,

pes

tici

de,

and

seed

s.C

ost

and

inves

tmen

tm

easu

red

indollars

.R

isk

aver

sion

ranges

0to

1,

wher

e0=

most

risk

lovin

gand

1=

most

risk

aver

se.

72


Tab

le27

:IV

Ex-A

nte

Lan

dby

Ris

kA

ver

sion

(1)

(2)

(3)

(4)

(5)

Ow

nla

nd

Ren

ted

land

Share

crop

land

Tota

lla

nd

Any

Cult

.

Tre

atm

ent

-0.0

32

0.0

44

-0.0

14∗

-0.0

10

-0.0

07

(0.0

26)

(0.0

36)

(0.0

07)

(0.0

51)

(0.0

53)

Ris

kA

ver

sion

X0.0

33

0.0

55

0.0

15

0.1

15∗∗

0.0

66

Tre

atm

ent

(0.0

31)

(0.0

44)

(0.0

10)

(0.0

57)

(0.0

55)

Ris

kA

ver

sion

0.0

94

0.0

40

-0.0

01

0.1

26

0.0

81

(0.0

65)

(0.0

51)

(0.0

10)

(0.0

82)

(0.0

73)

Contr

ols

Yes

Yes

Yes

Yes

Yes

Contr

ols

XR

isk

Aver

sion

Yes

Yes

Yes

Yes

Yes

Dis

tric

tF

EY

esY

esY

esY

esY

es

Mea

nD

ep.

Var

0.1

30.1

80.0

20.3

30.4

4O

bse

rvati

ons

5736

5732

5735

5731

5737

p-v

alu

eT

reat

+R

isk

XT

reat

0.9

49

0.0

00

0.9

64

0.0

01

0.0

31

Note

s:Sam

ple

incl

udes

all

obse

rvati

ons

from

both

trea

tmen

tand

contr

ol

gro

ups.

Tre

atm

ent

isin

stru

men

ted

usi

ng

firs

tyea

rel

igib

ilit

yin

tera

cted

by

yea

r.D

ata

isp

oole

dfr

om

both

the

2016

and

2017

Am

an

seaso

n.

Contr

ols

are

incl

uded

for

pre

cisi

on,

and

are

com

pri

sed

of

base

line

mea

sure

sof

tota

lla

nd

owned

,house

hold

size

,and

the

age

and

educa

tion

of

the

hea

dof

house

hold

.Sta

ndard

erro

rscl

ust

ered

at

bra

nch

level

.L

and

ism

easu

red

inacr

es.

Tota

lla

nd

isth

esu

mof

own

land,

rente

dla

nd,

and

share

cropp

edla

nd.

Any

Cult

.is

an

indic

ato

rfo

rw

het

her

or

not

ahouse

hold

pla

nte

dany

crops

duri

ng

the

seaso

n.

Ris

kav

ersi

on

was

mea

sure

dat

base

line

and

ranges

0to

1,

wher

e0=

most

risk

lovin

gand

1=

most

risk

aver

se.

73


Tab

le28

:IV

Ex-A

nte

Inp

uts

by

Ris

kA

ver

sion

(1)

(2)

(3)

(4)

(5)

Fer

t.A

pplied

Pes

t.A

pplied

Cost

See

ds

per

acr

eIn

put

Cost

per

Acr

eN

on-A

gIn

ves

t

Tre

atm

ent

-2.6

2-0

.05

0.6

2-0

.55

-13.8

0(9

.41)

(0.4

0)

(1.7

5)

(4.9

8)

(25.6

4)

Ris

kA

ver

sion

X17.5

20.5

5-1

.01

4.7

629.3

4T

reatm

ent

(13.1

8)

(0.5

6)

(2.3

0)

(6.5

5)

(33.2

7)

Ris

kA

ver

sion

-12.6

2-0

.95

-4.2

2-2

0.2

0∗∗

11.3

9(2

2.7

1)

(0.7

0)

(3.7

4)

(10.0

4)

(36.6

6)

Contr

ols

Yes

Yes

Yes

Yes

Yes

Contr

ols

XR

isk

Aver

sion

Yes

Yes

Yes

Yes

Yes

Dis

tric

tF

EY

esY

esY

esY

esY

es

Mea

nD

ep.

Var

140.0

81.5

616.8

366.6

152.6

7O

bse

rvati

ons

2550

2473

2423

2352

5737

p-v

alu

eT

reat

+R

isk

XT

reat

0.0

51

0.0

54

0.7

22

0.1

57

0.1

72

Note

s:Sam

ple

incl

udes

all

obse

rvati

ons

from

both

trea

tmen

tand

contr

ol

gro

ups.

Tre

atm

ent

isin

stru

men

ted

usi

ng

firs

tyea

rel

igib

ilit

yin

tera

cted

by

yea

r.D

ata

isp

oole

dfr

om

both

the

2016

and

2017

Am

an

seaso

n.

Contr

ols

are

incl

uded

for

pre

cisi

on,

and

are

com

pri

sed

of

base

line

mea

sure

sof

tota

lla

nd

owned

,house

hold

size

,and

the

age

and

educa

tion

of

the

hea

dof

house

hold

.Sta

ndard

erro

rscl

ust

ered

at

bra

nch

level

.F

erti

lize

rand

pes

tici

de

are

mea

sure

din

kg

/L

per

acr

e.In

put

cost

per

acr

eis

the

sum

of

the

cost

of

fert

iliz

er,

pes

tici

de,

and

seed

s.C

ost

and

inves

tmen

tm

easu

red

indollars

.R

isk

aver

sion

ranges

0to

1,

wher

e0=

most

risk

lovin

gand

1=

most

risk

aver

se.

74


Table 29: Ex-Post Outcomes 2016


Treatment 0.013 -0.004 128.861∗∗ -0.118(0.048) (0.050) (55.976) (0.114)

Flood X Treatment 0.144∗ -0.090 -142.596∗ 0.310∗

(0.074) (0.077) (80.684) (0.170)

Flood -0.094 0.058 -20.675 0.037(0.076) (0.077) (60.773) (0.142)





Notes: Sample includes only eligible BRAC members from both treatment and control groups. Data is from onlythe 2016 Aman season. Controls are included for precision, and are comprised of baseline measures of total landowned, household size, and the age and education of the head of household. Week interviewed fixed effects areincluded for the log consumption regression due to the presence of holidays over the course of the survey period thatchanged standard consumption patterns. Standard errors clustered at branch level. Income is measured in dollars.Flood is an indicator that equals one if flooding occurred and the Emergency Loan was activated.

Table 30: IV Ex-Post Outcomes


Treatment 0.040 -0.027 110.893∗∗ -0.155(0.053) (0.053) (46.853) (0.121)

Flood X Treatment 0.066 0.014 -100.623∗ 0.513∗∗∗

(0.062) (0.065) (51.432) (0.144)

Flood -0.019 0.029 -3.086 -0.130(0.047) (0.049) (31.068) (0.104)





Notes: Sample includes only eligible BRAC members from both treatment and control groups. Treatment isinstrumented using first year eligibility interacted by year. Data is pooled from both the 2016 and 2017 Amanseason. Controls are included for precision, and are comprised of baseline measures of total land owned, householdsize, and the age and education of the head of household. Week interviewed fixed effects are included for the logconsumption regression due to the presence of holidays over the course of the survey period that changed standardconsumption patterns. Standard errors clustered at branch level. Income is measured in dollars. Flood is anindicator that equals one if flooding occurred and the Emergency Loan was activated.

75


Table 31: Ex-Post Outcomes with out Flood Controls


Treatment 0.080∗∗ -0.019 47.896∗ 0.118(0.031) (0.029) (28.093) (0.076)




Mean Dep. Var 5.93 10.77 275.22 1.51Observations 4743 4531 4745 4745

Notes: Sample includes only eligible BRAC members from both treatment and control groups. Data is pooledfrom both the 2016 and 2017 Aman season. Controls are included for precision, and are comprised of baselinemeasures of total land owned, household size, and the age and education of the head of household. Weekinterviewed fixed effects are included for the log consumption regression due to the presence of holidays over thecourse of the survey period that changed standard consumption patterns. Standard errors clustered at branch level.Income is measured in dollars.

76

Credit Lines as Insurance: Evidence from Bangladeshbarrett.dyson.cornell.edu/NEUDC/paper_447.pdf ·...

Documents

Transcript of Credit Lines as Insurance: Evidence from Bangladeshbarrett.dyson.cornell.edu/NEUDC/paper_447.pdf ·...